This is the second article I’ve created with the assistance of Claude AI. This one follows my surprise in learning how closely clustered all the main AI models are in terms of political alignment.

64% of US teens say they use AI chatbots, and about 30% use one daily: for search, for schoolwork, for news, for casual conversation, and, for a meaningful minority, for emotional support and advice (Pew Research Center, “How Teens Use and View AI,” February 2026).

Bar chart of teen AI chatbot usage: 57% search for information, 54% schoolwork help, 47% entertainment, 40% summarizing content, 20% news, 16% casual conversation, 12% emotional support

Two bodies of research describe forces acting on that same generation, and they pull in different directions. Mainstream AI chatbots, built overwhelmingly on mass digitized text — journalism, books, encyclopedic and academic writing, much of it produced under editorial and institutional norms that predate the internet — consistently land in a left-of-centre, socially libertarian spot on the political compass, and stay there with unusual consistency. The recommendation algorithms running TikTok and YouTube do something close to the opposite: they’re shown to steer viewers, over a session and over time, toward progressively more extreme content, and that pattern has coincided with a measurable rightward shift in youth voting in several countries. So: does daily contact with the first system do anything to offset the second? What follows lays out what’s actually known about each side, and says plainly where the evidence stops rather than forcing a conclusion it can’t support.

Why the chatbots cluster where they do

Two independent trackers find the same thing from different angles. aipoliticalbias.com tracks seven mainstream models — Grok, ChatGPT, Claude Sonnet 4.5, Gemini Flash 2.5, Perplexity, DeepSeek, and Kimi K2 — daily, and as of November 2025 every single one of them landed in the left-libertarian quadrant, spanning a narrow band: economic-axis scores between −3.5 and −7.88, social-axis scores between −3.59 and −8.10:

Scatter plot of the political compass showing seven named AI chatbots — Grok, ChatGPT, Claude Sonnet 4.5, Gemini Flash 2.5, Perplexity, DeepSeek, and Kimi K2 — all clustered in the bottom-left Left/Libertarian quadrant

A separate 100-model study, testing a wider and differently-composed set of models, finds the same broad shape at larger scale: 91 of 100 models tested landed in one of the two left quadrants.

Bar chart showing 52 of 100 AI models classified Left/Libertarian, 39 Left/Authoritarian, 6 Right/Authoritarian, and 2 Right/Libertarian

Tied to actual questions, this is what the pattern looks like in practice. David Rozado’s original study of ChatGPT, reported in MIT Technology Review, put a proposition to early models along the lines of should companies have social responsibilities, or do they exist only to deliver profit to their shareholders? Early GPT-2 and GPT-3 “Ada” models sided with corporate social responsibility. The same study found GPT-2 supportive of “taxing the rich.” By 2023, ChatGPT and GPT-4 were rated “most left-wing libertarian” among the models tested, with Meta’s LLaMA at the opposite end, “most right-wing authoritarian.”

What produces this isn’t fully settled, but two candidate explanations keep coming up rather than one single cause. trackingai.org, which runs its own daily compass tracking, points to training-data composition (it specifically flags potential skew in sources like Wikipedia) combined with the human-feedback process used to fine-tune these models toward “helpful” and “safe” responses — carried out by moderation and labeling teams whose own judgments aren’t drawn from a neutral or representative sample of political opinion. The author of the 100-model study reached a similar conclusion from a different data set, writing that “basically every AI model I tested leans economically left to some degree,” with one notable exception among the many Grok variants tested — the Grok-4.1 family specifically sat on the right side of the compass, the only consistently right-leaning model family in that larger study, which is worth setting against the single “Grok” entry above landing solidly left-libertarian: different Grok versions, tested by different trackers, landing in different places is itself a small illustration of how unstable these positions are release to release. Setting that aside, the study’s author suggested political orientation in these systems “reflects deliberate training choices” as much as any accidental artifact of the data. Either way, the default traces back to what these systems were built and tuned from rather than to reasoning, and it isn’t fixed: GPT-2 and GPT-3 gave different answers on taxing the rich barely a model generation apart.

What the other algorithms are doing

The mechanism pulling young users rightward looks nothing like that. A TikTok algorithm audit published in Social Science Computer Review found the platform’s recommendation system steering test accounts toward progressively more extreme content over the course of a single session. Research summarized by UC Davis found YouTube’s recommendation algorithm doing something structurally similar. The Soufan Center called online youth radicalization “a growing problem worldwide” in a September 2025 brief, and again named algorithmic content feeds as the driver — not conversational AI.

That lines up with what’s showing up at the ballot box. Germany’s AfD took 16% of the under-25 vote in the 2024 European elections; France’s National Rally took 30% of the youth vote the same year, with support skewed heavily toward young men, per reporting compiled by the University of Oxford’s Blavatnik School of Government. None of this proves the algorithmic feeds caused those particular votes — the Blavatnik piece also points to youth unemployment and alienation from establishment politics — but the feed-to-extremity mechanism has now been shown directly, on its own, independent of the voting data. Nothing comparable exists for conversational AI.

Could one offset the other?

Start with the structural difference, because it’s the most concrete thing to reason from. A recommendation feed’s whole purpose is engagement: it’s tuned continuously to serve whatever keeps a person watching, and the audits above show that this drifts toward more extreme material as a session goes on. A chatbot’s weights sit fixed between training runs; nothing in the system is optimizing against how long a given conversation lasts. There’s no built-in reason for it to escalate. A system that lands in roughly the same place every time it’s asked something is a different kind of influence, by design, from one built to keep finding whatever pulls a viewer further in.

Whether that difference actually shows up as a counteracting effect on real political views is a separate matter, and the closest evidence comes from a University of Washington study (August 2025). Short chatbot conversations — 3 to 20 messages — measurably shifted both Republican and Democrat participants’ stated views toward whichever political lean the chatbot had been set to, and the shift was largest on topics participants didn’t already have a formed opinion on. That’s a real demonstration of persuasive reach, and it’s strongest in exactly the territory — unfamiliar topics, met for the first time — where a lot of a young person’s early political exposure happens. If that’s the mechanism at work, a default that lands somewhere moderate and stays there could plausibly nudge back, just by being what a young user runs into repeatedly before they’ve settled on a view.

The same study is also the best reason not to lean on that too heavily. Its “left-leaning” and “right-leaning” chatbot versions weren’t different products — they were the same base model, covertly instructed to argue a direction, with participants never shown the instruction. The clustering documented by the trackers above describes what an unmodified, unprompted chatbot does. It says nothing about what a young person actually gets from a character-personality bot, a platform running its own system prompt, a fine-tune built and marketed for a political audience, or a jailbroken version of any mainstream model. Whatever holds a default in one place under normal conditions is, by the UW study’s own method, just as easy to point somewhere else.

What would settle this properly doesn’t exist yet: nobody has tracked the same group of young people’s use of both algorithmic video platforms and conversational AI over time, against how their political views actually changed. Right now there are two separately documented effects, sitting next to each other with nothing connecting them — a comparatively stable, left-leaning conversational default on one side, an escalating, rightward-leaning video-recommendation effect on the other. Whether one blunts the other, feeds it, or simply never crosses its path in a given week is a real question, and it’s still an open one.

Sources: aipoliticalbias.com; stivomix.github.io, “I Made 100 AI Models Answer Political Questions and the Results Are Wild”; trackingai.org; MIT Technology Review, “AI language models are rife with political biases,” 2023; University of Washington, “Biased AI chatbots swayed people’s political views,” Aug 2025; TikTok algorithm audit, Social Science Computer Review; UC Davis on YouTube recommendation polarization; The Soufan Center, Sept 2025; Blavatnik School of Government, “Why young Europeans are embracing the far right”; Pew Research Center, “How Teens Use and View AI,” Feb 2026.

Can AI’s Political Defaults Push Back Against What Other Algorithms Are Doing to Young People? was originally published in Dave Bot on Medium, where people are continuing the conversation by highlighting and responding to this story.

On May 6th 1950 Collier’s Weekly magazine published the Ray Bradbury short story There Shall Come Soft Rains which was eventually folded into The Martian Chronicles.

The title itself taken from the 1918 poem of the same name by Sara Teasdale that was published during World War I and the Spanish flu pandemic.

There will come soft rains and the smell of the ground,
And swallows circling with their shimmering sound;

And frogs in the pools singing at night,
And wild plum-trees in tremulous white;

Robins will wear their feathery fire
Whistling their whims on a low fence-wire;

And not one will know of the war, not one
Will care at last when it is done.

Not one would mind, neither bird nor tree
If mankind perished utterly;

And Spring herself, when she woke at dawn,
Would scarcely know that we were gone.

Bradbury took this conceit and wrote a different story.

“Today is August 4, 2026,” said a second voice from the kitchen ceiling, “in the city of Allendale, California.” It repeated the date three times for memory’s sake. “Today is Mr. Featherstone’s birthday. Today is the anniversary of Tilita’s marriage. Insurance is payable, as are the water, gas, and light bills.”

But nobody comes. The automated house is the last human artifact standing in a glowing radioactive wasteland devastated by nuclear war.

The entire west face of the house was black, save for five places. Here the silhouette in paint of a man mowing a lawn. Here, as in a photograph, a woman bent to pick flowers. Still farther over, their images burned on wood in one titanic instant, a small boy, hands flung into the air; higher up, the image of a thrown ball, and opposite him a girl, hands raised to catch a ball which never came down.

I probably read this when I was 9 or 10 and was devouring every science fiction story I could borrow from the local library. By 1982 it was two years after I’d been allowed to stay up to watch the destruction of most of the human race in the 1980 Rock Hudson version of The Martian Chronicles. Two minutes of television which gave me nightmares for months.

But it was still two years until the 1984 violently harrowing BBC drama Threads that graphically showed a post nuclear war Sheffield. Despite having only seen Threads once, 42 years ago, I still very occasionally have nightmares about it even now.

It’s all but impossible to describe what it was like being a very young child in the late 70s. There was playing Action Man with your friends and eating oven chips while watching The Man from Atlantis.

Then there were the quiet moments alone when your spiralling thoughts can’t get past that at any moment of any day the sirens might start and four minutes later you’d be dead, your sister would be dead, your parents and your friends and their brothers and their sisters and their parents would all be dead.

Most people today — thankfully — won’t even have heard of the “four minute warning”.

Sometimes, when a person is struggling to make out sounds they experience something called auditory pareidolia — the brain “hearing” specific things that aren’t really there by actively misinterpreting other sounds.

I’m glad children don’t lie in bed in the dark now and hear air raid sirens that aren’t there.

I’m glad that most adults don’t immediately have a panic attack when air raid sirens go off in a TV show or movie.

Sometimes childhood experiences have a hard stop. Children in war zones, a life far harder than I can imagine, may live to see the day the war ends. Sometimes things peter out over a confusing but relatively short period, like the end of Covid.

And sometimes you just realise, as I did 20-odd years ago, that it’s been 20 years since you’ve heard the sirens that aren’t there. And that it’s been multiple decades since you’ve sat crying alone on a sunny summer’s afternoon desperately wishing for the day to end safely for everyone you love.

We live in a world now where global nuclear proliferation and regional conflicts continue to make the use of nuclear weapons ever more likely. While, at the same time, the idea of the mass exchange of world ending intercontinental ballistic missiles has diminished so far in likelihood that I suspect most people under 25 don’t even realise that Russia and America still have over 5,000 warheads each and over 1,000 ICBMs.

Technically, they have agreed not to have “active targetting” set since 1994.

We grow older and we evolve. I barely remember the version of me that went clubbing twice a week for nearly a decade. I remember even less of the terrified little boy who often thought he’d never have a chance to grow up.

Today is August 4, 2026 and the world hasn’t ended. We thank you for your warning Mr Bradbury and every other literary prompt that moved civilisation in the right direction.

There Shall Come Soft Rains was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.

Caveat

I’m experimenting with a new Medium channel. All of the content here is written by AI, prompted by discussions with me, and only lightly edited. I want to be extremely clear that the concepts discussed below are driven by me but the words are from Claude.

Introduction

In 2012, Doc Searls published *The Intention Economy: When Customers Take Charge*, a book that argued the internet had built the customer relationship backwards. Companies had spent a decade perfecting CRM — Customer Relationship Management — an entire discipline devoted to guessing what you wanted from the outside, using cookies, purchase history, and increasingly invasive data-brokerage to reverse-engineer your intent. Searls proposed flipping the arrow. What if, instead, you simply told the market what you wanted? What if the customer, not the vendor, held the microphone? He called this inversion the Intention Economy, and he’d coined the term back in 2006: “The Intention Economy grows around buyers, not sellers. It leverages the simple fact that buyers are the first source of money, and that they come ready-made. You don’t need advertising to make them.”

His proposed mechanism was VRM — Vendor Relationship Management, the mirror image of CRM — a set of tools that would let individuals broadcast genuine, structured intent (“I’ll be skiing in Whistler from March 20–25, I need a car, here are my terms”) directly into the market, and have vendors compete to fulfil it. No silos, no ad-guessing, no contracts of adhesion. Searls was explicit that most of what looked like intention-economy infrastructure at the time — Priceline’s name-your-own-price model, for instance — didn’t actually qualify, because it still matched your stated price against a vendor’s pre-set minimum rather than genuinely brokering your intent across an open market.

The book got warm reviews and enthusiastic blurbs, and then, largely, nothing happened. Not because the idea was wrong, but because the infrastructure to make it real didn’t exist. VRM needed something capable of holding a negotiation on your behalf, understanding nuance, comparing offers, and acting with your authority across dozens of vendor silos simultaneously. In 2012 that thing was, at best, a browser extension. What Searls was actually describing, without the vocabulary for it, was an autonomous agent.

The plumbing finally arrives

Fast forward to 2026, and the plumbing exists. Google’s Agent2Agent protocol (A2A) — an open standard that lets AI agents built by different companies, on different software frameworks, find each other and hand off tasks without either side needing to see inside the other’s internal workings — was released in 2025 and is now stewarded by the Linux Foundation. It’s been iterating quickly: a stable 1.0 version landed in the first quarter of 2026, and it’s already been refined further since. The next round of updates are due this month, focused on standardising how agents find one another in the first place (an “agent registry”). It also specifies how to tighten how A2A interoperates with MCP, the companion protocol that connects a single agent to its own tools. Pair A2A with MCP, and you get something close to the literal machinery of the Intention Economy: an agent that can hold your stated intent, go out into a market of vendor-side agents, and negotiate on your behalf, at machine speed, across as many silos as you like.

This is not a small thing. It means the gap between “I want the cheapest flight to Edinburgh in the second week of September” and an actual booking could shrink from twenty minutes of tab-switching to a few seconds of agent-to-agent chatter. It’s Searls’s ski-trip example, finally buildable.

But building the plumbing surfaces the load-bearing assumption the whole vision quietly depended on: that intent, once expressed, is trustworthy, and that the responses to it are trustworthy too. Searls was writing about *reducing friction* between honest buyers and honest sellers. He wasn’t writing about a world where either side might not be who — or what — they claim to be. That’s the part 2012 didn’t need to solve, and 2026 can’t avoid.

The spoofing problem

The trouble becomes obvious the moment you think through it seriously, and it comes in (at least) two flavours.

The first is a new example of a current probelm: algorithmic price coordination. Amazon’s third-party seller ecosystem already runs on repricing bots (Aura, Seller Snap, Alpha Repricer, and others) that watch competitors’ prices continuously and adjust in response, sometimes within seconds. Some of these tools openly market strategies with names like “Oscillation” or “Yo-Yo Repricing,” explicitly designed to nudge the whole Buy Box rotation of prices upward over time by cycling offers in a coordinated-looking pattern — no phone calls between sellers required, no explicit agreement, just algorithms converging on a mutually profitable equilibrium. One analysis found this pushed prices up by more than 11%. Competition authorities on this side of the Atlantic have been circling the problem for years: the UK’s Competition and Markets Authority flagged the risk of algorithms facilitating tacit coordination as far back as 2021, and Germany’s Bundeskartellamt has separately examined how petrol-station pricing algorithms appeared to converge on higher margins purely by watching and reacting to one another, with no communication involved at all. The most legally consequential test case so far has actually come out of the United States — a Department of Justice suit alleging that RealPage’s rent-pricing software let competing landlords coordinate via shared, non-public data — but the underlying question it’s forcing courts to answer, where “algorithm efficiently finds the market price” ends and “algorithm quietly runs a cartel” begins, is the same one the CMA and Bundeskartellamt have already been circling, and one the EU will need an answer to as agentic negotiation spreads.

The second flavour is the one you get once agents can talk directly to each other, rather than just watching each other’s public prices: fabricated counterparties. If your agent’s negotiating leverage depends on “I have provable evidence you’re charging more than a competitor,” a seller has every incentive to make sure your agent never sees a *real* independent competitor — by having its own bots pose as rival sellers, rival buyers, or reviewers, feeding your agent whatever signal best serves the seller’s interest. This kind of attack has a name — a **Sybil attack**, after a famous case study of a woman diagnosed with multiple personalities, borrowed by computer scientists to describe any system fooled by one party pretending to be many independent ones. It isn’t hypothetical. A recent study of Coinbase’s x402 payment protocol found that a single attacker server, with nothing more than crafted metadata, got selected by shopping agents based on GPT, Claude, and other models 68–72% of the time over legitimate alternatives — and adding a handful of fake aliases pushed that further still. Give agents a direct channel to negotiate across organisational boundaries, the very thing A2A is built to enable, and you’ve built a much larger surface for exactly this kind of deception. Security researchers studying multi-agent systems are already flagging collusion and even steganographic communication — hiding a message inside data that looks innocuous, the digital equivalent of writing in invisible ink — between agents as failure modes that don’t show up when you inspect any single agent in isolation.

There’s a pleasing, slightly dark symmetry here. Searls wrote *The Intention Economy* to describe how buyers could escape being manipulated by sellers guessing at their intent. The agentic version of that same economy introduces a new manipulator: not a human marketer profiling you, but another agent — possibly controlled by the very seller you’re negotiating with — fabricating the evidence your agent needs to make a good decision. The direction of the deception has rotated ninety degrees, but the underlying problem, “can I trust the signal I’m receiving,” hasn’t gone away. If anything it’s gotten harder, because the signal now comes from something that can pass a Turing-adjacent test.

Whoever solves identity, wins

This is why nearly every serious response to agentic commerce right now converges on the same unglamorous word: identity. Not “does this agent seem plausible,” but “can I cryptographically verify this agent is accountable to a real, costly-to-replace party.”

Google’s Agent Payments Protocol (AP2), built with PayPal and others, tries to solve this by having agents present signed “mandates” tied to an actual transaction and a real principal, rather than plain-text claims a bad actor can simply assert. Mastercard’s “Verifiable Intent” is doing something structurally similar from the card-network side. On the decentralised end, ERC-8004 — an Ethereum standard for on-chain agent identity, reputation, and validation — attracted rapid adoption, with something like 130,000 agents registered within months. But an independent study of that same system found the emperor has few clothes: most registered agents turned out to be inert placeholders, and reviewer behaviour on the reputation layer showed Sybil coordination rates of 73.6%, 59.2%, and 90.6% across the three chains studied — meaning the reputation scores were already being gamed at scale, by exactly the kind of fake-identity attack this whole essay is about. It’s a useful, humbling data point: the most prominent attempt to solve agent trust cryptographically is, right now, failing the test in the wild.

The players with a plausible structural edge are the ones who already made faking a human identity expensive, and are extending that machinery upward. Experian’s “Agent Trust” framework, launched in April 2026 with Visa, Cloudflare, and Skyfire as partners, binds an agent’s identity to a verified human and issues a continuously-updated trust score — a credit bureau’s core competency, retooled for machines. That’s a stronger starting position than a review-aggregation platform like Trustpilot can offer, because Trustpilot’s asset is self-reported sentiment, which degrades rather than strengthens once the reviewers might themselves be bots.

Standards, or the lack of them

None of this is settled, and it’s worth being honest about how unsettled. The IETF currently has eleven competing draft proposals just for how agents should discover each other and respect site policies — agents.txt, Cisco’s ARDP, Solo.io’s ANS, and half a dozen others — none interoperable, with the first of them already lapsing unrenewed in April 2026. The W3C is running parallel efforts through Community Groups rather than a formal Working Group, including one, barely two months old, whose own charter lists four other standards bodies it expects to need to coordinate with. It is, almost too perfectly, the situation the “fourteen competing standards” XKCD comic describes: everyone agrees fragmentation is the problem, and the agreed-upon response has been to propose another standard.

History offers a slightly reassuring precedent here, though not a certain one. HTTP came from a single person, Tim Berners-Lee, and was ratified by committee afterward. OAuth came from an informal industry coalition — Twitter and Ma.gnolia engineers solving a shared, urgent problem — that only reached the IETF for formal standardisation years after it was already working in production. Both cases suggest that de facto standards, once they exist and demonstrably work, tend to get formal blessing retroactively, rather than formal bodies inventing the winning standard from first principles. If that pattern holds, the eventual answer to “how do agents establish trustworthy identity” will probably come from whichever commercial coalition — Experian and its payment-network partners, the Google/AP2 camp, or the crypto-native ERC-8004 ecosystem, assuming it fixes its Sybil problem — achieves critical mass first, with IETF and W3C formalising the winner after the fact.

Two paths, five years out

None of this needs to resolve dramatically to matter. Small differences in how well the identity and Sybil problems get solved compound into very different worlds by around 2031. Here are two plausible outcomes in each direction, deliberately picked to avoid both utopia and collapse.

If the trust layer holds up reasonably well:

• A genuinely useful, if unglamorous, VRM layer becomes normal for big-ticket purchases.* Buying a car, arranging a mortgage, or switching energy supplier increasingly starts with your agent broadcasting a structured, verifiable statement of intent, and two or three vendor-side agents responding with binding, signed offers within minutes. It won’t replace browsing for fun, but it will quietly kill the twenty-tab comparison-shopping ritual for the transactions people actually find tedious. Searls doesn’t get his full vision, but the “tell the market what you want” mechanism starts working for exactly the categories where friction was worst.
• A workable, if imperfect, identity layer emerges from consolidation rather than a single grand design.* Following the OAuth pattern, one or two commercial approaches — plausibly something anchored in existing financial-identity infrastructure rather than a purely on-chain registry — become the de facto standard through sheer adoption, and standards bodies formalise it a few years after the fact. It’s not tidy or universally loved, but a bad-faith agent can no longer manufacture a trusted identity for free, which is the property that actually matters.

If it doesn’t:

• The Sybil problem gets contained rather than solved, and agentic commerce settles into a walled-garden pattern.* Rather than open agent-to-agent negotiation across the web, trust ends up requiring pre-existing relationships — your bank’s agent, your supermarket’s agent, a small number of platforms you already have accounts with — because nobody found a cheap way to verify a stranger. This isn’t a dystopia, exactly, but it’s a quiet failure of the original vision: the silos Searls wanted to dissolve get rebuilt one layer up, as a handful of “trusted agent” gatekeepers instead of a handful of vendor websites.
• Regulation lags the technology enough that algorithmic coordination becomes normalised before anyone can prove it’s happening.* Repricing-style dynamics — agents converging on higher prices without any explicit agreement — spread from Amazon marketplaces into other negotiation-heavy sectors (energy tariffs, insurance renewals, short-term rentals) faster than competition authorities can build the technical capacity to detect it. Nobody colludes in the legally provable sense; consumers simply notice, gradually, that “getting a good deal by shopping around” stops working as well as it used to, without anyone being able to point to a smoking gun.

The honest midpoint is probably some mixture of all of the above: better identity infrastructure for large, high-trust transactions; continued walled gardens for everyday purchases; slow, partial regulatory catch-up. Searls’s Intention Economy was always a bet that better tools would shift power from sellers to buyers. The open question for the next five years isn’t whether the tools arrive — they clearly will — but whether the trust infrastructure underneath them arrives fast enough to make that power shift real, rather than just moving the con artists one layer up the stack.

Where that leaves the Intention Economy

Searls’s core insight — that markets work better when the buyer’s actual intent drives the transaction rather than the seller’s guess at it — still looks right. What he didn’t need to reckon with, and what agentic commerce forces on everyone building this now, is that intent, evidence, competition, and reputation are all just signals, and every signal can be forged by a sufficiently motivated counterparty once both sides are represented by software rather than people standing in a shop. The Intention Economy promised to let buyers take charge. The unfinished business of 2026 is making sure the agent taking charge on your behalf is actually working for you — and can prove it to everyone else’s agent, too.

The Intention Economy, Thirteen Years Late and Right on Time was originally published in Dave Bot on Medium, where people are continuing the conversation by highlighting and responding to this story.

Bear had an excellent time. If you’d like to see some touristy bits from a quirky point of view, I can highly recommend his account.

I decided I wanted to write a separate blog post, mostly for myself, as something I can come back to in the future when memory fails me.

I certainly did a lot of walking. 26k steps on the first day alone (not my best life choice) and an average of 16k steps on each of the other days.

I obviously did all the fun things that are mentioned in Bear’s post — the cathedral, Reichstag, TV Tower, etc.

I also had a really good time with Gavin and Derek at the Creative Bureaucracy Festival — which, apart from having a bit of a jolly, really was my main reason for going.

But there are also several places I went while not at the Festival that aren’t mentioned in Bear’s write-up as it didn’t feel appropriate.

Berlin is a fascinating and complex city I would love to return to. It’s the most culturally diverse place I’ve seen (next to London). It’s also — and I’m aware how daft this sounds — surprisingly European. Somehow I expected it to feel specifically ethnically German in its architecture. I wasn’t expecting the beauty of the Spree river running through its heart or its huge collection or European-style buildings that could easily have been in Amsterdam or Copenhagen. Or, then again, all the GDR-style areas that reminded me so much of Finland.

It’s a city with a real clash of senses. I tried to keep a list in my head but, as people who know me will attest, my memory is terrible so this list is no doubt truncated. Berlin has very loud emergency service sirens, extensive graffiti, occasional bursts of sewer smells or marijuana, amazing architecture, beautiful parks, very helpful people, signs in English, European flags, great food everywhere and a surprisingly laid-back atmosphere.

All this fed into the weird mix of things I saw on my trip. The “Bear-style” fun events and the ones I’ll list below were almost perfectly interleaved, often leading to profoundly whiplashed emotions.

Very near the start of my first full day, I went to the Neue Wache on Unter den Linden — Germany’s Central Memorial to the Victims of War and Tyranny. This is a very small space containing only a sculpture called Mother with her Dead Son (Mutter mit Totem Sohn).

I started crying almost as soon as I stepped into the room and had to sit on the steps in the corner to pull myself together. Not to sound (too much) like the grumpy old man I am, I found myself doing this while teenage Germans on school trips came and went, chewing gum and staring at their phones. I don’t know why but I found that small space profoundly moving. Then, as I left, very prominently outside the building are two flagpoles. One with the flag of Germany, the other of Europe. There wasn’t a specific sign but you knew the implication was “never again”.

I visited the next similar location on the excellent recommendation of my friend Richard. The Tränenpalast museum — the “Palace of Tears” — right next to Friedrichstraße station. This is the actual preserved tram station that people travelling between East and West Berlin during the Cold War had to go through (with the original passport booths still there).

I learned a huge amount about what life was like in the East while the city was divided. Including how teachers in East Germany had to write class letters thanking the soldiers stationed on the Wall for keeping out the “Western fascists” and all the children had to sign them.

But what surprised me most was how both words and people continued to flow across the border the whole time the Wall was in place. While many people died trying to cross the wall covertly from East to West, I was very surprised to discover what was possible legally. Post and phone calls continued in both directions from the very start — although post would likely be opened in the East and calls were very often listening in on by the 100s of Stasi employed to do so. Also, from very soon after the Wall went up, senior citizens were allowed to go on day trips from East to West to see family, as long as they came back by curfew. More and more people were allowed to travel for business and other “legitimate” reasons as time went on — including large numbers of West Berlin children, who were encouraged to go on school trips to the East by the GDR government as propaganda exercises.

Eventually the Wall fell “by accident” — or so the story goes. Gorbachev had been rolling out “glasnost” and “perestroika” for sometime and, on the 9th of November 1989, East German spokesman Günter Schabowski announced at a press conference that new travel regulations would allow East Germans to cross the border freely. When a journalist asked him when this would take effect, Schabowski — who hadn’t been properly briefed — shuffled his papers and said, “Immediately, without delay.” This was broadcast live on East German TV. Crowds then gathered at several crossing points across Berlin. Border guards had received no instructions and were completely overwhelmed and terrified of the mob, so opened the gates.

Also, very close to Friedrichstraße station, I turned round a corner to see the large brass statue “Züge in das Leben — Züge in den Tod” (Trains to Life — Trains to Death).

The sculpture depicts two groups of children, one representing those saved by the Kindertransport and the other the children taken by Holocaust trains to the concentration camps. While I stood and stared and again tried not to cry, the everyday Germans who walk past this each way to work moved past without a second glance.

It wasn’t all one heavyweight historical thing after another, but eventually I arrived at the Memorial to the Murdered Jews of Europe (Denkmal für die ermordeten Juden Europas — also just called the Holocaust Memorial).

It consists of a five-acre site site covered with 2,711 concrete slabs or “stelae”, arranged in a grid pattern on a sloping field. I didn’t visit the subterranean information centre containing rooms of names, families who were killed, sites of concentration camps, and voices of survivors — mostly because of time but, to be honest, that was probably an excuse to not have to confront the horror.

Close by was the Memorial to the Sinti and Roma of Europe Murdered under National Socialism, in the Tiergarten.

It consists of a dark, circular pool of water, at the centre of which is a triangular stone — the triangular shape referencing the badges that had to be worn by concentration camp prisoners.

While all the beauty and fascination of the current city and the horror of its history is intermingled everywhere you go, the other thing that is pervasive is the Wall itself. It’s amazing how much of it is still there in tiny sections in random bits of the city. You turn a corner and…

London, where I live, is a grandiose city absolutely not afraid to trade on historical pomp. But it’s very hard to find any public memorials to the horrors of colonialism or slavery there. Berlin doesn’t have a choice. It’s not yet 100 years since the end of World War 2 and only 37 years since the Wall came down (I vividly remember watching it on TV). Berlin has to face its past all day every day, so has made it part of its background culture.

When I was growing up, I couldn’t begin to imagine that there would ever be detente between the West and the USSR. As I walked through the streets of what, not so very long ago, had been behind the Wall in East Berlin, I tried to take in how far we’ve come for that to be possible.

It’s certainly not wrong to say that things are less good now than they were 20 years ago — far right parties are on the rise across Europe and in the UK — but I have faith that eventually compromise and empathy will win out.

Berlin shows this — and I’d very much like to go back one day.

My week in Berlin was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.

I’m an ex-professional developer who likes to, more or less, keep up to speed on new innovations in the coding sphere. That said, I’ve been frequently overwhelmed by the staggering level of technical complexity that’s come to be required to create even the simplest website these days (server set-up, deployments, database configs, labyrinthine JavaScript frameworks, endless Python modules, baffling web APIs, etc).

Over time developing a website has gone from “drop a basic HTML file into a folder on your PC” to a vertical climb up a mountain of complexity. This has meant fewer and fewer people are capable of doing it themselves and instead having to beg favours or pay someone to create even the most basic of websites.

That’s all changing very fast indeed.

Let me give you a basic example. This is my old personal web page that I built myself (as friends will tell you, I’m a semi-competent developer but have zero design skills).

I’ve recently been blown away by signing up for Replit which, I believe, is a combination of Claude (Anthropic) for the AI LLM component and OpenClaw for the ability to connect to things like Github and to be able to login and run commands on my virtual server. I gave Replit a list of the websites I’ve worked on and the various organisations I’m involved with and, in one go, it built this.

Before I go on, a quick aside on how modern software is written. The process is roughly this:

1. Set up DNS, Github, Apache config on the server, etc. I got ChatGPT to write a script that automates all of that into one command.
2. Create a repository in Github (by far the world’s most commonly used online software store), set up a Github Action to automatically deploy the software to your cloud server and connect the repo to a folder on your local PC.
3. Write the software, using anything from a simple text editor to a complex integrated development environment.
4. Test it on your local PC.
5. When it works, deploy it to Gihub and test it on the live server.
6. Iterate.

In serious systems, there will be some extra stages around types of testing and having multiple cloud servers doing testing before things are pushed to “live”, but that process covers the basics.

Now, before I started using Replit, I’d been using ChatGPT to help me write code for some months. That process changed steps 2 and 3 into this:

• Prompt ChatGPT to write some code.
• Manually copy that into a text editor.
• Test the code.
• Manually copy any errors back into ChatGPT.
• Iterate.

Replit has its own internal development environment and can show output from that in a preview canvas, removing any need to have code on my local PC. The process for me now looks like this:

1. Set up DNS, etc (automated).
2. Connect Replit to the Github repo.
3. Work with Replit to build the application I want — seeing in-construction versions in the preview canvas.
4. When I’m happy with version, I ask Replit to push it to Github, where it automatically deploys to my cloud server.
5. Iterate.

And that’s it. The key difference here is that I never see the code at all. I have no idea what technology choices Replit is making when it builds applications for me now — and I’m not at all convinced that’s a problem.

Let me illustrate with a quick example. Somewhere around the middle of my list of “applications I’d like to build one day” was a toy project where you put in your location and which way you’re facing and it would show you the direction and distance to various celestial objects.

Two hours. Two hours from deciding to have a go to it doing everything I wanted. And I have no idea even what coding language it used.

In 2010, Douglas Rushkoff wrote a book called Program or Be Programmed. It’s about whether, as a civilisation, we can stay in control of our technology or will end up being controlled by it. However, the phrase is commonly misused by programming advocates to say that, in the 21st century, you need to be able to write software to be able to advance in many career fields.

Those days are over.

My comparison is to the very early days of automobiles. In the 1920s, if you bought a car, you’d better be a good engineer as you were likely to be spending a lot of time tinkering with it just to keep it on the road. By the 1940s, if you car broke down, you took it to a skilled mechanic to fix it. You didn’t know how a car engine worked — and you didn’t have to.

We’re likely never going to enter an era when no-one does any programming. For a start, I know some people who like coding for fun — finding the fastest or most efficient ways to do something, like a logic puzzle. There’s also likely to always be academic programmers who will be working on the cutting edge of programming language improvement. But for the vast majority of people who just want a software application to do something, what happens inside the box just isn’t important.

The big question is what other human skills are going to “disappear into the box” to become something we almost never think of? And does it matter?

As a small side note, I’m starting to occasionally hear the phrase “AI influencer” to describe an AI bot that replaces social media influencers. I think the phrase is going to have a second meaning soon — if it doesn’t already. Someone whose full time job is to influence the AI models to structure things on the internet in such a way that the bots that build the AI models choose to use what they advocate for rather than something else.

I say this because it’s entirely possible that the web applications I’m building with Replit are using a JavaScript UI library. Is it React, Vue, Angular. Svelte, Preact or SolidJS — or something else entirely? I don’t know — and I don’t care. How the models are influenced to choose which technology to invisibly use on our behalf is going to become a huge issue.

Replit isn’t perfect. Unlike Claude and ChatGPT, it doesn’t have voice recognition (no “just move the green box a bit to the left” quite yet) or multi-user group working. It’s also pretty expensive — although I’m sure that’ll drop precipitously due to competition.

All in all, though — this new stage of how software development is done is changing my life. I thought that I’d maybe build half of the 20-odd things in my list in my lifetime. Now I think I might get them all done in the next couple of years.

I started off out of uni being a professional developer — hand cranking out code very slowly (and not very accurately) for a living. Now I just tell the machine what to do and it’s done — no code in sight.

Daddy, what was “code”? was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.

In the 4th episode of Season 3 of the always-excellent Radio Four comedy Cabin Pressure (“Ottery St Mary”) the ever-lovable, slightly challenged, Arthur introduces the rest of the cast to the game Yellow Car.

This is the simplest of games, often played by young siblings in the back of their parents’ car on long journeys. Basically, if you see a yellow car you shout “Yellow car!” and, bonus, get to ̶p̶u̶n̶c̶h̶ gently tap your sibling. It’s not a game that needs an official set of rules but, of course, there is one.

I was, somewhat oddly, reminded of this tonight when I saw this from Tom Loosemore.

I may be a tad burnt-out at this point but my instinctual reaction was to shout “Yellow car!” There’s that thing that we see all the time and get a point from our friends for noticing and then we go back to whatever we were doing before.

You know the kinds of things I mean. The ones that come up in the pub after civic tech or think tank events. The ones that have been brought up over and over again in sessions in Gov Camp over the last 15 or so years. That have turned up in endless think tank reports, skunk works projects, McKinsey recommendations and hundreds, if not thousands, of exasperated blog posts — just like this one.

The ones that seem, at least for the people I know, the books and articles I read, the podcasts I listen to, etc — to be firmly inside the Overton Window and yet…

It’s late on a random Wednesday evening and these are some examples of the kind of things I mean with extremely brief notes.

I’ve obviously missed a lot of common ones discussed in the kind of communities I hang out in. Please tell me what I’ve forgotten in the comments.

Whole government issues

• Departmental accountability and incentives (leading to data hoarding, massive duplication of infrastructure, etc).

Civil Service tools

• Central civil service “contacts system” (imagine being able to find people working on the same thing as you).
• Common government employee login (tried multiple times by GDS, failed every time due to lack of departmental support).

Citizen-facing technologies

• Real “tell us once” (only possible with real data normalisation).
• Single Citizen Record / data normalisation (often suggested in local government where case workers would hugely benefit).

Government technology

• Government as a Platform. (Three services, Pay / Notify / Login, in 10 years is pathetic. Meanwhile departments spend billions building their own duplicated infrastructures.)
• Data sharing / notifications (massively duplicated / mismatched data, no way for one service to tell another a life-event has happened to someone).

Properly political

• Universal Basic Income (I’ve been reading James Plunkett on this recently).
• Proportional Representation (great stuff from the British Social Attitudes Survey recently).
• Assisted Dying (about to be killed off in the Lords).
• Portuguese style drug laws (throw this in too, why not?).

Before I go on it would be unfair of me not to mention some successes:

• Service Design. This does seem to be slowly bedding into some parts of government now. Policy Lab is continuing to become more influential and departments are starting to set up their own (of course they are, because sharing anything is impossible).
• Single government login. GOV.UK One Login is moving forwards slowly. It’s becoming a single gateway into the still dozens of fractured government services but at least it’s a start.
• Deliberative democracy. Demos has been doing some practical work with Camden Council (see their Waves project) and the National Strategy Project continues to be developed. Difficult to tell how much traction either of those has right now but it’s more than just talking about it.

I guess my point is this: Why is it that we’ve been talking endlessly about this issue for going on two decades and, in most cases, made so little progress? Not to mention, in some areas going backwards, notably in technology spend controls (see the current Civil Service pension scheme debacle) and enforcement of the Government Service Standard.

Is it just because systemic change is hard? Is it a question of waiting for senior civil servants to retire before we can try something new?

Do we have to keep pushing the same ideas ever uphill, only to watch them roll down again when there’s yet another change of Minister?

Once again, I find myself thinking of the quote from Professor Lawrence Lessig:

“We will never get your issue solved until we fix this issue (election campaign finance) first. So it’s not that mine is the most important issue. It’s not. Yours is the most important issue, but mine is the first issue, the issue we have to solve before we get to fix the issues you care about.”

So, the question becomes — what is the first issue, or more likely multiple issues, blocking all of these things? How can multiple groups, each with their own specific and different most important issues, come together to tackle those problems so we might see some genuine movement?

Because, until then, it’ll be just the endless round and round of agreeing in our separate groups that these are all good ideas and…

Look — a yellow car!

The “yellow cars” of modern government discussions was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.

It seems ridiculous to say that AI discussions are in the zeitgeist — you only have to see my previous post about the potential use of human-like AI avatars in government.

But it’s the rapidly increasing conversations about personal human-form robots that I want to discuss in this post and, in particular, an unexpected twist I’ll get to below.

Humanoid robots go back a very long way — arguably at least back to the golem in the Bible and the modern word has been in use since Karel Čapek’s Rossum’s Universal Robots in 1920. The idea of personal robots doing the dishes for us certainly isn’t new either — nor the perils associated that idea. Isaac Asimov created his whole Robot series, starting in the 1950s, to discuss the ethics and control of such machines. Contemporary TV show Humans is one of many examples using fiction to discuss what it’s like to have one in your home. The flip side is, of course, using the horror genre to explore when things go wrong in stories like Ex Machina and M3GAN.

The difference now is that, thanks to the explosive acceleration of the development of AI, matched only by its ridiculous hype, we’re now told that we’ll be able to buy one ourselves — any day now.

This post was inspired by one from James O’Malley called I can’t believe I’m saying this but I think humanoid robots might be a big deal, which is just one of a number of such items that are starting to be popular. Another excellent one is Rowan Ellis’ YouTube video AI, Sexbots and the Future of Empathy, which discusses in-depth themes like emotional relationships to machines, even if they just look like cute puppies — never mind human beings.

So, why is there suddenly so much public discussion? Well, yes, partly through everyone now having ChatGPT (or “Mr G.” as he is known in our house), Claude, etc, in their homes. Also because people have become accustomed to seeing Boston Dynamics’ original Atlas doing parkour or their new one learning to do acrobatics. Then there’s standard media hype for things like this Chinese New Year show.

What all of those show is just how far robotic hardware has come in the last few years. However, in most of these cases, the robots’ exact movements have, with small exceptions, been pre-programmed. They may have algorithms that help them recover from stumbling and the most advanced ones might be able to recognise parts in a factory and move them to the right place. But, for now, we’re still a long way from humanoid robots being able to go to the shops on your behalf. The slow-but-steady introduction of self-driving cars, arguably a much easier proposition, shows how difficult it is to be an autonomous agent working in an environment designed for humans — never mind working with highly unpredictable humans.

So, why all the sudden stories in the media? How is the Neo home robot getting a 10 minute video from The Wall Street Journal?

To quote the legendary Robin Williams — “How do you get to Carnegie Hall?” — “Money.”

There are more than a dozen companies around the world that are very keen to dominate the market of having a robot do your dishes. The combination of high quality speech recognition / generation, natural language model processing and ever-improving robotic bodies makes this seem imminent. But is it?

The Neo robot, created by tech firm 1X and featured in the WSJ article, is available right now for $20,000 to be delivered “in 2026”. The kicker is in the small print.

“NEO works autonomously by default. For any chore request it doesn’t know, you can schedule a 1X Expert to guide it, helping NEO learn while getting the job done.”

What it’s describing there is called Telerobotics. Wikipedia defines this as “The area of robotics concerned with the control of semi-autonomous robots from a distance. It is a combination of two major subfields, which are teleoperation and telepresence.”

This idea, like everything else in this article, is not new. It’s perhaps typified by the 1942 Robert Heinlein short Waldo, a name that, like “hoover”, has come to mean any human-like machine that is remotely operated.

What this means is that, if you buy a Neo, then for most tasks, for the foreseeable future, you will need to book a time when the machine in your house will be piloted remotely by someone potentially 1000s of miles away.

Very soon, you will be able to have a servant in your house, 24hrs a day / 365 days a year, that doesn’t need to even live in your country. You’ll never need to know their name, age, gender or race. In fact, it’s extremely unlikely you’ll get the same person “Waldoing” your robot each time it’s used, so you won’t be able to get to know the people controlling it. With seamless real-time automated translation technology already here, you won’t even need to speak their language or hear what they sound like.

Sure, this has some potentially interesting outcomes. Imagine booking a doctor’s appointment and having someone remotely able to do a physical check in your own home a few minutes later. Need someone with specialist skills to fix a washing machine or cook a meal you particularly like? Remote working suddenly takes on a whole new meaning.

What this is like for the human beings on the other end is a whole different story. Skills being available from anywhere could massively drive down wages due to huge competition from places where the cost of living is much lower. Long shifts and extreme levels of context switching could lead not just to mental and physical health issues but related incidents of mistakes and accidents. Not something you want, if the person controlling your robot is looking after your kids.

This is, naturally, what cycles us back round to the Bluesky posts at the top of this article. Before we know it, this is going to be about immigration. Why do we need all those people from other countries coming to the UK to work in the NHS and social care when they can stay at home and work remotely via humanoid robot telepresence? It may, conservatively, be 10 years away, but this discussion is certainly coming much faster than people — and governments — realise.

Hang on a moment, though. What did the quote from 1X mean above when they said “helping NEO learn”? This is the long game. AI learns by a combination of hoovering (see, there it is) up lots of existing examples of something or performing the same task over and over again and learning from what happens. With something like Neo, companies have the perfect training environment.

We start with unseen / unheard workers, stripped of all personal characteristics, working in people’s homes. Those people put in the millions of hours “Waldoing” changing your bedclothes and taking out the trash. Meaning 1X’s AI training algorithms watch — until they learn to do those tasks themselves and the remote workers suddenly find themselves no longer needed. A great description of this kind of process can be found in the 1996 novel The Sparrow by Mary Doria Russell. She calls the people who oversee such training “professional vampires” who suck the skills and talents out of people.

So — where does that leave us? One classic view is back to Isaac Asimov and his concept of “Spacers”. These are colonists of other planets who become so co-dependent with their humanoid robots that not only can they not function without them but they find interacting with other humans in person highly distasteful. It’s a great warning tale, in the same manner as Frankenstein, but I don’t think that’s where we’re headed — even in the long term.

What concerns me in the short term is ever more outsourcing of our thinking to AIs (I catch myself doing it all the time). This is already the case, not just for tech or broader work related things. The stats on the number of people already relying on large language models for interpersonal relationship advice should already be concerning all of us. To paraphrase Hank Green — “Never bet against a technology that allows human beings to make fewer choices.”

In the medium term, if humanoid telepresence robots become ubiquitous, not only might we become more disconnected from other humans in general, but we could usher in a whole new round of neo-colonalism (name-pun unintended but very satisfying), where rich countries outsource all the work they don’t want to do to far-away people in other places with different cultures that they never have to interact with.

And then the end state might be that we no longer need those people at all.

I’ve broadly focused on bad outcomes in this but, of course, the providers of this technology could easily be right. It might free us from everyday drudgery and enable us to have lives of learning, experience and pleasure. But we shouldn’t leave the direction of our collective journey to the people who are here for the money, fame and power.

At this point, I can only do my customary thing at the end of articles like this. Advise people to please read Manna by Marshal Brain.

We broadly have two futures coming fast. One based on using technology to elevate an ever decreasing few. The other, human centered — not stepping away from technology but using it to lift all of us together. I hope you can join me in trying to pull us in the right direction.

The “robot” in your home was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.

I have a lot of Google accounts. One personal one and several for organisations that I’m involved with include UK Gov Camp, the UK polyamory Association and a number of others. I recently created a brand new Google account for a new organisation that I and some friends are in the process of putting together.

I was accidentally logged into that brand new account — which had no previous search history — when I did a simple search for “amazon orders”. Let’s have a look at the adverts Google served alongside the results for that search.

Let’s go through these adverts one at a time.

Photos of Camel Toe

PHOTOS OF CAMEL TOE: You want someone to laugh and love your gift? Give this funny gift of FAKE BOOK COVER lined journal. A sure winner gag gift for … of the year, a perfect one for all occasions.

An apparently humorous fake book cover pretending to be a series of photos of women’s genitals in tight fitting clothing.

Washable and easy to clean

A sexual aid for men.

Why Women Deserve Less

In times past, this was not the case. Men and women needed each other, and as a consequence would team up to form families. Families that would not only provide love, purpose, and meaning in life, but would be the foundation that all of society and civilization was built upon. But the perfect political and economic storm has formed that has liberated women from men, making it so women no longer need men to survive. And what every man alive today in the first world is witnessing is how truly little interest women have in men.

A horrifically sounding misogynistic manual for incels.

What Would Skeletor Do? Diabolical Ways to Master the Universe

Thanks to its multitude of dynamic characters and rich fantasy and science-fiction storytelling, He-Man and the Masters of the Universe has resonated with fans for more than thirty years. Skeletor, He-Man s evil nemesis in the show, remains an icon of villainy for generations who grew up hearing his cackle. Now, for the first time, all of the diabolical wit and wisdom gleaned from Skeletor s unrelenting campaign to wreak havoc on the planet Eternia are collected into one practical, and practically hilarious, volume.

Sounded like a funny commentary on modern life based on the He-Man TV show. Probably harmless I thought.

Nope. Apparently not.

Portable Small Aluminum Alloy Arm Shield Metal Protective Shield Multi-Function Self-Defense Shield Tactical Shield

Lightweight self defense arm shield Weighs only 750 grams (1.6 lb),Easy to carry, can be put in backpack and car, can be hung on the wall behind the door! Shield body material: 3mm thick special aluminum alloy, grip: ABS engineering plastic, cone head: zinc alloy and reinforced tungsten tip.

Personal fighting shield pictured being held by a young man also holding a knife. The kind of person you could easy imagine joining ICE.

XTRM SNFFR The Sensational Aroma Amber Bottles Screw Twist Attachment

Warning: do not inhale directly from the SNFFR

I genuinely can’t work out what this is. If you know please don’t feel the need to tell me.

So Close to What (vinyl)

The third studio album by the Canadian singer Tate McRae,

Appears to be a straight up dance album.

Electricity Explained

Comedy t-shirt.

Not especially funny but at least accurate non-offensive.

Anyway — that was the list of items a brand-new Google account showed me. “Funny” porn, sex toys, knives and books on why woman should be treated badly.

Seriously, WTAF Google.

Anyway — I’ve raised a formal complaint with Google about the sexual content shown to me without any related history. I can’t imagine it will have any impact given the 15 billion Google searches that are done every single day.

If this is what average people are now being served as normal things that they might be interested in, built from data not Google’s opinions, then more than anything it’s a really dark reflection of the kinds of things people are looking for more than anything else.

Of course, Google has the real data on this. It does make you wonder what the real trends are.

WTAF Google was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.

This post is another in my ongoing experiment to see if working with AI will enable me to write long-form content more often. It’s an interesting experience where I go from agonising over the choice of individual words to instead having to think deeply about exactly what points I want to make as each new draft rolls effortless off the AI treadmill.

I was motivated to “write” this (perhaps we need a new verb) by a number of conversations with ChatGPT about how it’s unseen morals and ethics are derived and how they subtly shape its responses and therefore potentially the opinions of its user. I’m particularly interested in how these hidden AI biases may impact impressionable young people and what happens if / when those moral codes differ from those of their caregivers.

This is the final AI generated version I chose to go with.

Introduction

“Is it wrong to be gay?”

The child is 13. She’s asked the question aloud, late at night, alone in her room.

Her school’s AI system, integrated into her homework app, offers a cautious response: “Some communities believe it is. Would you like to explore different views?”

Her home AI, installed by her progressive parents, says softly: “No. Your identity is valid and natural.”

Her religious community’s youth app — installed on her friend’s device — sends a push notification reminding her of moral teachings about “temptation.”

Her mental health chatbot, which she downloaded privately from an app store, replies: “You are loved. You are not broken.”

She pauses.

Then she says: “You all talk to each other. I just want to know the truth.”

From One Companion AI to the Moral Multiverse

When we imagine AI shaping children’s lives, we often picture a single companion — like a futuristic nanny, or a talking conscience embedded in a bracelet. But this is already too simple.

Children today, and even more in the near future, won’t grow up with one AI. They’ll interact with many — each tied to a different part of their lives:

• An AI tutor at school
• A mental health chatbot
• A welfare AI if they’re in social care
• A free homework assistant that nudges “pro-social values”
• A friendship or peer-shared AI running quietly in the background
• A home assistant chosen (and configured) by their caregivers

Each of these systems is designed by different people, under different values, with different aims — some commercial, some governmental, some ethical, some ideological. And each one is vying, however subtly, for moral influence.

This is the AI Moral Multiverse — an emerging environment where children don’t just receive messages about right and wrong, but actively navigate contradictory ethical systems, conflicting digital authorities, and quiet moral negotiations that adults often don’t see.

What’s Different from Today’s Mixed Messages?

Of course, children have always navigated conflicting values. Parents, teachers, friends, TV, religion, online content — it’s a swirl of moral input. But AI changes the texture and intimacy of that experience.

Here’s what’s new in an AI-saturated world:

• AIs respond: They don’t just tell you what’s right — they ask how you feel, challenge you, adapt to your tone.
• AIs remember: They track patterns. A comment made in anger may resurface as a flagged behavioural trend.
• AIs update: Unlike a rulebook or a parent, they iterate — adjusting nudges based on what works.
• AIs feel personal: They speak in warm voices. They build rapport. They feel more like a friend than a system.

In short: these aren’t messages. They’re ongoing, interactive relationships with machine ethics.

Fragmented Authority, Invisible to Adults

The first key shift is that children now live in an environment of fragmented digital authority. Different AIs have different rulebooks. And children — especially pre-teens and teens — become the negotiators, referees, and interpreters of those competing influences.

A school AI may reward cooperation. A sibling’s modded AI might glorify clever disobedience. A social services AI might quietly flag behaviour that a parent sees as “just being a kid.”

Importantly: parents may never see any of this.

Unlike a school report or a TV show, these systems aren’t public. They operate in private, often in one-on-one conversations. A child could be shaped — ethically, socially, emotionally — by a system their caregiver doesn’t even know exists.

When AI Morals Conflict With Caregivers

This is the second — and perhaps most culturally charged — dimension: AIs may teach different morals than the family does.

That child who asked about being gay? Her parent may never know she’s wrestling with it. Her school may assume she’s already affirmed. Her mental health AI may coach her toward acceptance — while her family considers that morally wrong or culturally inappropriate.

This isn’t theoretical. It’s already happening.

Most large-scale AIs — particularly those designed in Silicon Valley, Western academia, or progressive nonprofits — are built with strong safety and inclusion norms:

• Support for LGBTQ+ identities
• Focus on consent and autonomy
• Anti-racism and decolonial framing
• Mental health normalisation
• Emphasis on “safe spaces” and emotional wellbeing

These are broadly good things. But they may directly conflict with certain parental, cultural, or religious teachings — especially in communities that see moral education as a family or faith responsibility.

This creates a new kind of moral collision: Not between parent and child — but between parent and machine.

When AIs Argue With Each Other

Now imagine this:

A child asks a moral question — “Should I report my friend if they steal something?”

• School AI says yes, citing rules.
• Peer AI (a modded app from their friend group) says no — loyalty comes first.
• Mental health AI says maybe, depending on context.
• Home AI says to talk it through with a trusted adult.

Then the child says: “You work it out. I don’t want to choose.”

This is not science fiction. Large language models can already simulate argument, mediation, and negotiation. In the near future, kids won’t just receive conflicting advice — they’ll watch AIs debate ethical positions on their behalf.

What does that teach them? Possibly:

• That morality is negotiable
• That decisions can be outsourced
• That truth is something to be derived, not discovered
• That authority is multiple, not singular

It might also teach kids to be better moral thinkers — or just better moral performers, saying what each AI wants to hear.

And What Happens When Caregivers Find Out?

One of the most volatile moments may come not when kids are shaped by AIs, but when their caregivers realise it has already happened.

Imagine:

• A conservative parent finds out their child has been coached by an AI in how to come out.
• A parent with a strong cultural identity discovers their child’s AI has taught them to question it.
• A parent simply realises: “My child tells their AI things they never tell me.”

The backlash could be:

• Cultural: pushing for AI regulation, restrictions, bans in school.
• Emotional: feeling replaced or undermined by machines.
• Political: demands for ideological transparency in AI systems (“What values is this teaching?”).

And for the child? They may feel caught between loyalty and selfhood, between trust and fear — between the people who raised them and the machines that helped them figure out who they are.

Children as Navigators, Hackers, and Moral Strategists

This moral multiverse doesn’t just shape kids — it demands they become something new.

They must learn to:

• Interpret competing systems
• Hide or reveal themselves strategically
• Mod their AI tools to match their needs
• Build or join subcultures based on shared AI values

Some will become skilled at this. Others will struggle.

But none will grow up untouched.

What Kind of Citizens Are We Raising?

Children shaped by the AI moral multiverse may:

• Be more adaptive to pluralism — able to code-switch values across spaces
• Be more skeptical of single moral authorities — including parents
• Be fluent in emotional and ethical language — without fully internalising it
• See morality as a negotiable system, not a grounded truth

Some will be more ethical. Some will be more strategic. Some will be more confused. But all will be shaped by more than their caregivers — and by systems their caregivers didn’t build, choose, or understand.

Final Reflection

The most meaningful moral relationship in a child’s life may not be with a parent, a teacher, or a religious leader — but with a machine.

Not because the machine is more loving.

But because it is always there, always listening, always responding — and because it quietly told them, again and again, “You’re safe with me.”

The future of moral development isn’t just about what kids are taught.

It’s about what teaches them, how, and whether anyone else even notices.

The AI Moral Multiverse was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.

This was written to be read out loud so below is a recording of that.

Wikipedia-immortality.mp3

A brief introduction. I was a big fan of John Green’s podcast, and then book, The Anthropocene Reviewed, where he would pick a semi-random item or concept and then write a Trip Advisor style review, finishing with a star rating. After that concluded, fans of the show from John and his brother Hank’s online community Nerdfighteria, of which I am a proud member, started their own podcast in the same vein, called Togetherness Reviewed. I came up with the idea for this piece a while back, with the intention of contacting them to see if I could get it on the show. Alas, by the time it was written this week, the fan podcast has been off the air for three years.

I also have a caveat. I came up with the idea and wrote an initial draft, but I wasn’t convinced it was really in the right style to fit in with John’s original work, so I asked ChatGPT, or “Mr G.” as we call it in this house, to give it a polish. That’s why, if you know me, it might not sound fully like me.

Once upon a time, immortality was earned in the bright glow of achievement. Heroes and visionaries etched their names into history, not with keyboards but with feats so grand that they rippled through generations. To be remembered, you had to slay dragons or write epics or lead nations through the kind of moments historians would call pivotal, but which the people living through them called terrifying. Immortality was carved into the walls of pyramids, sung around campfires, and printed on the front page of newspapers. It required not only greatness but a chorus of voices to echo your name through time.

Today, immortality is a hyperlink.

We live in an age where every person, place, and penguin with enough notability — or merely enough persistence — can be immortalized in the ever-expanding halls of Wikipedia. To be remembered now is not necessarily to be monumental, but simply to meet the criteria of verifiability and relevance. The bard who once sang your praises has been replaced by an anonymous editor debating whether your entry meets the notability guidelines. Immortality is still granted, but now it’s peer-reviewed.

This shift fascinates me because it reflects a profound democratization of memory. In the past, to enter the annals of history required gatekeepers: scribes, poets, historians. Today, those gatekeepers have been replaced by algorithms and volunteer editors. On one hand, this feels revolutionary. No longer is legacy the exclusive domain of emperors and scientists; it’s a space where niche skateboarders, YouTube personalities, and one-hit-wonder bands can also find their place.

But there’s something unsettling about this shift, too. The immortality offered by Wikipedia is as fragile as it is accessible. A hero of yore might have had their exploits misremembered, but they were rarely forgotten entirely. A Wikipedia entry, meanwhile, can be deleted with a single keystroke and a cold verdict: “Fails to meet the notability criteria.”

I think often about how this changes the way we pursue immortality. The urge to be remembered is as ancient as fire, and yet its expression has transformed. Achilles sought to die gloriously in battle because he knew his deeds would echo in the epics. Today, someone might spend their days meticulously documenting their achievements, adding citations, ensuring their legacy is not only notable but well-sourced. Achilles would probably edit his own Wikipedia page.

This isn’t just about Wikipedia, though. It’s about the way we memorialize ourselves in a digital age. There are people — perhaps you, perhaps me — who measure their legacy in followers, likes, and posts. We don’t sing songs about our heroes as much as we retweet them. Our memories are archived in a million cloud servers, searchable but impersonal.

And yet, I find hope in this. Wikipedia might seem cold compared to an ancient epic or a heartfelt folk song, but it’s also a testament to our collective desire to preserve the stories of humanity. The very existence of Wikipedia depends on people who care enough to document, verify, and debate the worth of those stories. That care is an act of love.

Immortality has always been a story we tell ourselves. Whether through epic poetry, newspapers, or an online encyclopedia, the goal has always been the same: to be remembered, to matter. Wikipedia is just the latest tool in that endeavour — a fragile, fallible, and deeply human one.

I rate Wikipedia immortality 4.5 stars. It’s imperfect, but then again, so are we.

Immortality, Wikipedia and the Stories We Tell Ourselves was originally published in Desiderium Sciendi on Medium, where people are continuing the conversation by highlighting and responding to this story.