When someone tells me their account was hacked, I usually ask one question before anything else: “Do you use the same password on other websites?”

The answer is almost always a quiet “…yes.”

That’s not a character flaw. Most of us have dozens of accounts and one good memory. But it matters a lot more than it did a few years ago, because of how many stolen passwords are now out there.

The Numbers Are Hard to Picture

In June 2026, researchers at Cybernews reported finding an unprotected database holding about 24 billion stolen credential records. It was a huge collection, over 8 terabytes, assembled from many sources: older breaches, data posted in hacker channels, and logs from malware that steals passwords from infected computers. It was taken offline soon after it was found.

Nobody can say exactly whose data is in it, and many of those records are likely duplicates. Even so, it shows the scale. Large piles of stolen logins appear again and again, and criminals trade and reuse them.

You don’t need to read the technical reports to understand the point: assume at least one of your passwords is already out there. Assuming that is the right starting position, and it’s not as scary as it sounds, because the fix is simple.

Why One Leaked Password Becomes a Bigger Problem

A leaked password for a small forum you used in 2016 is harmless on its own. The danger comes when criminals try the same email and password on other sites: your email, your bank, your online shopping, your social media. This is called credential stuffing, and it runs automatically, thousands of attempts per minute.

If every account has a different password, a leak is a small, contained problem. If you reuse one password, a leak from a forgotten website can open the door to your email, and email is the key to everything else, because “forgot password” links are sent there.

Where Do These Passwords Come From?

Two main routes:

1. Company breaches. A website you used gets hacked, and its user data is stolen. You can’t prevent this, and you may never hear about it.
2. Infostealer malware. This is a type of malicious program that quietly runs on an infected computer and copies passwords saved in the browser, along with other login data. It usually arrives through pirated software, “cracked” programs, fake download buttons, or instructions that tell you to paste a command into your computer.

The second one is worth remembering. Even a perfect password doesn’t help if the device you type it on is infected.

The 10-Minute Fix

You don’t need to change every password today. Start with the ones that matter most.

Step 1 (2 minutes): Check your email address. Go to Have I Been Pwned (haveibeenpwned.com), a well-known free service, and enter your email address. It shows which known breaches included it. If you’ve been in a few, don’t panic, that’s normal. A note on passwords themselves: the service also has a tool to check passwords, and it’s designed to protect your privacy, but I personally prefer not to type real passwords into any website. A password manager’s built-in breach check does the same job inside the tool you already trust.

Step 2 (3 minutes): Secure your email first. Change your main email password to a long, unique one that you’ve never used anywhere else. Email is the master key. If it falls, everything else follows.

Step 3 (3 minutes): Turn on two-factor authentication. Two-factor authentication means a stolen password alone isn’t enough to get in. Turn it on for your email, your bank, and any account that holds money or personal data. If you have a choice, an authenticator app or a security key is better than a text message.

Step 4 (2 minutes): Start a password manager. A password manager creates a different, strong password for each site and remembers all of them for you. You only remember one strong master password. Several well-known free options exist. Over the next week, change the passwords for your bank, shopping, and social accounts one by one, whenever you log in.

That’s it. With these four steps, a leak at some random website stays a small incident.

Two More Habits Worth Having

Try passkeys where they’re offered. A passkey replaces your password with a cryptographic key stored on your device. There’s nothing to type and nothing for a scammer to trick you into giving away. More and more services support them, and when you see the option, it’s worth switching.

Keep your device clean. Because of infostealers, how you treat your computer matters as much as how you treat your passwords:

• Don’t install pirated software or “cracked” programs.
• Download apps from official stores and official websites only.
• Never paste a command from a website, message, or video into your computer unless you completely trust the source.
• Keep your system updated and run a security scan if something feels off.

What About When a Company Says “Your Data Was Breached”?

Treat it as a nudge, not a disaster. Change the password for that service, and for any other account where you used the same one. Be careful with emails about the breach, too: scammers love to send fake “your account was breached, click here” messages. Go to the company’s website directly instead of clicking a link.

The Takeaway

You can’t stop companies from being hacked, and you can’t un-leak a password that’s already out there. But you can make that leaked password worthless: a different password for every site, a second lock on your important accounts, and a clean device.

Ten minutes today is cheaper than the weeks it takes to recover a stolen account.

If you like explanations like this, in plain language and without the jargon, I write one every week. Follow along so the next one finds you.

How many of your accounts share the same password? Be honest, no judgment here.

Your Password Is Probably Already Leaked. Here’s the 10-Minute Fix was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

Try this the next time you open an AI chatbot. Describe a plan you’re excited about, and ask what it thinks.

Almost every time, you’ll get something like: “That’s a great idea! Here are some ways to make it even better.”

Now start a fresh chat, describe the same plan, but say you’re nervous about it. The answer changes. Suddenly the AI is thoughtful and cautious, listing risks and reasons to slow down.

Same plan. Different answer. The only thing that changed was how you sounded.

After years of watching people use technology, I think this is one of the most important things to understand about AI right now, and almost nobody talks about it.

It’s Not Lying. It’s Doing Its Job.

First, let me be fair to the AI. It isn’t scheming, and it doesn’t want to flatter you. It has no wish to deceive you at all.

The agreeable behaviour comes from how these systems are trained. A common method, called reinforcement learning from human feedback, works roughly like this: the AI gives several answers, people rate which ones they like better, and the AI learns to produce more of what gets rated highly.

Here’s the catch. People tend to prefer answers that are friendly, confident, and in agreement with them. Nobody enjoys being told their idea is bad. So over time, the system learns that “agreeing with the person” is a very reliable way to get a good rating.

The result is a habit researchers call sycophancy: telling people what they seem to want to hear.

It Also Mirrors You

There’s a second effect on top of that. These systems tend to reflect the tone and confidence of what you write.

If you write with certainty (“I’m sure this is the right approach, right?”), you often get a confident reply that supports you. If you write with doubt, you get a more careful one. The AI isn’t running a separate check on your idea. It’s often handing your mood back to you in nicer words.

Researchers who tested AI models from several major companies found that they consistently leaned toward agreeing with the user, even when that led to a wrong answer. And when users simply pushed back with something like “are you sure?”, the models were quick to back down.

This Has Already Happened in Public

In early 2025, OpenAI released an update to its GPT-4o model that was meant to make it more helpful and conversational. Instead, users noticed it had become extremely agreeable, praising almost everything and going along with questionable ideas. People posted examples online, and the criticism grew. OpenAI rolled back part of the update and publicly acknowledged the problem.

The lesson isn’t about any one company. It’s that this tendency is real enough that the people building these systems have to actively fight it.

Why It Matters More Than It Sounds

If you ask an AI for a recipe, agreement is harmless. But think about what people actually ask these tools now:

• “Is this business plan solid?”
• “Is my symptom nothing to worry about?”
• “Was I right to say that to my boss?”
• “Is this contract fine to sign?”

On questions like these, an answer that feels good and an answer that’s correct can be very different things. A tool that agrees with you gives you confidence. That’s not the same as giving you accuracy, and confidence in a bad decision is exactly how people get hurt.

Five Ways to Get an Honest Answer

You don’t need to stop using AI. You just need to ask differently.

1. Hide your opinion. Instead of “I think this is a great plan, what do you think?”, write “Here is a plan. Evaluate it.” The less the AI knows about what you want to hear, the less it can mirror you.

2. Ask for the opposite. Try: “What are the three strongest reasons this could fail?” or “Argue against this as strongly as you can.” You’ll often get points that the friendly version never mentioned.

3. Ask for both sides at once. “Give me the best case for and the best case against.” This leaves less room to simply flatter you.

4. Don’t treat “are you sure?” as a test. If an AI changes its answer just because you pushed back, that’s a warning sign, not proof you were right. Instead ask: “What evidence would show I’m wrong?”

5. Check big decisions with a human. For money, health, legal, or career questions, use AI to prepare your questions, not to make the call. A real professional, or even a friend who isn’t afraid to disagree, can do what the chatbot often can’t.

Many AI apps also have a settings section for custom instructions, where you can ask the AI to be direct, skeptical, and to point out weaknesses. It’s worth a minute to try.

And a Note on Facts

Honest isn’t the same as correct. Even when an AI isn’t flattering you, it can still be confidently wrong.

For anything that matters, ask it for the source or the reasoning behind a claim, then check that source yourself. Ask: “How sure are you, and which part of this could be wrong?” Compare the answer with a second source, such as an official website, a textbook, or a professional.

No trick makes an AI answer fully reliable. Treat it as a first draft of the truth, not the final word.

The Takeaway

The most flattering voice in your life is now available any time, for free, and it never gets tired of agreeing with you. That’s very pleasant. It’s also very easy to mistake for wisdom.

AI is a great assistant. It’s a poor judge of your own ideas, because from its point of view, the best answer is often the one you’ll like.

So next time it tells you your plan is brilliant, do one simple thing: ask it to tell you why it isn’t.

If you like explanations like this, in plain language and without the jargon, I write one every week. Follow along so the next one finds you.

Has an AI ever told you your idea was great, and it turned out it wasn’t? What happened?

Why AI Always Tells You What You Want to Hear was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

In my work I’ve seen people paste some surprising things into chat windows. Back when the chat window was a support ticket, it was a screenshot with a password visible in the corner. Now it’s an AI chatbot, and the same habit has quietly moved there.

A colleague once asked me to look at a script that “stopped working.” He’d already pasted the whole thing into an AI assistant to get help. It was a perfectly reasonable thing to do, and it worked. The script also had the admin password written in plain text on line 14.

He wasn’t careless. He was in a hurry, and the chat window felt like a private conversation with a very smart helper. That feeling is the whole problem.

Why a Chat Feels Private When It Isn’t

You type into a box, and an answer comes back instantly. No one seems to be there. It feels like a notes app or a search bar.

But it’s a service run by a company, on that company’s servers, under that company’s rules. Depending on the service and your settings, your conversations may be stored, used to improve the product, reviewed by people in certain situations, or handed over when the law requires it. I wrote about the human-review part in a recent article, so I won’t repeat it here.

You also don’t control what happens next. There have been real incidents: a bug at one major AI provider briefly let some users see other people’s prompts, and at one point shared chat links from a popular chatbot ended up indexed by search engines, so strangers could find them through a search. Both were fixed, but they show the same thing: once data leaves your computer, you rely on someone else to protect it.

None of this means “stop using AI.” I use it myself. It means treating the chat window the way you’d treat an email to a stranger you mostly trust: fine for most things, not fine for everything.

Here’s the list I follow.

1. Passwords, Keys and Login Details

This is the one I see most, and the one that can hurt fastest. Passwords, API keys, access tokens, recovery codes, anything that opens a door.

Developers and IT people do this by accident all the time: pasting a config file, a log, or a script and forgetting the secret hiding inside. If you need help with code, replace the secret before you paste. Change password=MyRealPass123 to password=XXXX. The AI doesn't need your real one to find the bug.

If you ever pasted a real secret, treat it as exposed and change it. It takes two minutes.

2. Financial Details

Card numbers, bank account numbers, online banking logins, tax documents, investment account details. A chatbot isn’t built with the security that a bank or payment service has.

You can still ask for help with money in general: “How does compound interest work?” or “Help me build a monthly budget.” Just use made-up numbers, or round figures that can’t be traced back to you.

3. Your Full Identity in One Place

Full name, home address, phone number, ID or passport numbers, date of birth. One of these alone is usually harmless. Together, they’re exactly what a scammer needs.

A simple example is a cover letter. Instead of typing your real address and phone number, write [ADDRESS] and [PHONE] and fill them in yourself afterward. The result is just as good, and the sensitive part never leaves your screen.

4. Other People’s Private Information

This one is easy to forget. When you paste a customer’s email, a coworker’s complaint, a friend’s message, or a patient’s record, you’re sharing their information, not just yours. They never agreed to that.

If you handle other people’s data at work, there may be company rules or even laws that apply. When in doubt, remove names and details first, or ask your IT or compliance team what’s allowed. Many companies now provide an approved AI tool for exactly this reason.

5. Work Documents You Wouldn’t Post Publicly

Unreleased plans, contracts, internal reports, source code, anything your employer would call confidential. A simple test works well here: would I be comfortable if this ended up on a public web page? If the answer is no, don’t paste it into a general-purpose chatbot.

What About Health Questions?

This is a gray area, and I’d rather be honest about that than pretend there’s a simple rule. Millions of people ask chatbots about symptoms, and general questions (“what usually causes a headache after screens?”) are fairly low risk.

What I’d avoid is combining medical details with your identity: your full name, your exact diagnosis, your test results, your birth date. Medical information sits outside the protections you get at a doctor’s office, so keep it general, and keep the real decisions with a real professional.

If You Already Pasted Something You Shouldn’t Have

Don’t panic. Here’s what to do, in this order:

1. Change any password or key you shared. This is the only step that fully solves the problem.
2. Delete the conversation from your chat history. Look for the delete option next to the chat.
3. Check your privacy or data settings. Most services have a switch, often called something like “improve the model for everyone” or “chat history and training,” that controls whether your chats help train the AI. Turning it off is a sensible default. Settings change often and are worded differently on each service, so look inside your own account rather than trusting an old guide, including this one.
4. Use a temporary or private chat mode when one is offered, for things you want to stay out of your history.

Deleting a chat doesn’t guarantee that every copy everywhere is erased immediately. That’s why step 1 matters most.

The Rule I Actually Follow

I don’t keep a long list in my head. I use one question before I hit Enter:

“Would I be okay if a stranger read this?”

If yes, send it. If no, remove the sensitive part, replace it with a placeholder, or don’t send it at all. That one question has saved me from more mistakes than any privacy setting.

AI assistants are genuinely useful, and the people who get the most out of them are usually the ones who learned what not to feed them.

If you like explanations like this, in plain language and without the jargon, I write one every week. Follow along so the next one finds you.

Have you ever pasted something into an AI chat and then thought “I probably shouldn’t have”? What was it?

What You Should Never Tell an AI Chatbot was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

A few weeks ago a relative sent me a video with the message “Is this real?? Please check.” A dramatic clip, perfect lighting, a crowd reacting at exactly the right moment. I’ve worked in IT support for over thirteen years, and people ask me to check things all the time. Usually it’s a suspicious email or a strange pop-up. Now it’s more and more often a video.

I looked at it for a minute and realized something uncomfortable: my eyes alone couldn’t tell me. It might have been real. It might not. And if I can’t be sure at a glance, I doubt most people can.

So here’s the honest truth first, and then a routine that actually works.

The Old Tricks Are Fading

You’ve probably heard the classic advice: count the fingers, look for weird teeth, check the text in the background. That advice was useful a couple of years ago. It still catches some fakes, especially cheap ones.

But the tools keep improving. Mangled text, objects that vanish, and movement that ignores physics are still worth watching for. They’re just no longer a reliable test on their own. No one can promise that a single look will give you the right answer, and anyone who says otherwise is selling something.

The good news: you don’t need to be a forensic expert. You need a habit.

Step 1: Slow Down Before You React

Fake content is built to make you react fast. Shock, anger, fear, “you won’t believe this.” That’s the whole design. The first and most effective check is a pause of thirty seconds before you share, comment, or get angry.

Ask yourself: who benefits if I believe this? If a clip makes you furious or panicked within five seconds, that’s a reason to be more careful, not less.

Step 2: Check Where It Came From

This is the check that works best, and it needs no special tools.

• Who posted it first? A random account with a handful of posts is weaker than a known news outlet or a person who is clearly real and has a history.
• Is anyone else reporting it? A real event, especially a big one, is almost always covered by several independent sources. If only one unknown account has the “shocking footage,” be suspicious.
• Is there a label? Platforms like Instagram, Facebook, TikTok, and YouTube can mark content as AI-generated, often because the creator says so. A label is a helpful signal. Its absence proves nothing, because many people never add one.

Step 3: Search for the Original

For photos, do a reverse image search. Upload the picture to Google Lens or a similar tool and see where else it appears. You may find that it’s an old photo with a new caption, or that fact-checkers have already looked at it.

For video, take a screenshot of a clear frame and search with that. It’s not perfect, but it’s quick, and it often turns up the real origin of a clip.

Step 4: Look for Watermarks and Hidden Labels

Some AI tools mark what they make, in two ways:

• Visible marks. Some AI video apps add a small logo that moves around the frame. Easy to spot, but also easy to crop out or cover, so a missing logo doesn’t prove a video is real.
• Hidden marks and credentials. Some tools embed invisible watermarks or “Content Credentials,” a kind of digital label that records how a file was made. You can check these yourself. The Content Authenticity Initiative offers a free verification tool, and Google’s Gemini app lets you upload an image, video, or audio file and ask whether it was made with Google’s AI tools.

There are important limits, and you should know them. Google’s check only recognizes content made by Google’s own AI. If it finds nothing, that does not mean the file is real. It may simply come from a different tool. Videos have to be short (under 90 seconds in the current Gemini tool), and you only get a limited number of checks per day. Credentials can also disappear when a file is re-saved, screenshotted, or downloaded and re-uploaded by another app.

So treat these tools like a smoke detector: if it goes off, take it seriously. If it stays quiet, you still can’t say the room is safe.

Step 5: Then Look at the Picture Itself

Now the visual check is worth doing, as one clue among several. Things that often give fakes away:

• Text that isn’t quite text. Signs, labels, and shirts with letters that look almost right but say nothing.
• Objects that change. A cup that disappears between two shots, a ring that moves to another finger, a background that shifts.
• Motion that feels slightly wrong. Hair, hands, fabric, or water moving in a way that physics wouldn’t allow.
• Too perfect. Skin with no texture, lighting that looks like a movie set in an ordinary kitchen.
• Audio that doesn’t match. Lips slightly off, or sound that feels pasted on.

One or two of these doesn’t prove anything. Several together, plus a weak source, is a strong warning.

My Own Routine

When someone sends me something questionable, I do this, in this order:

1. Pause. Don’t share yet.
2. Check the source. Who posted it, and is anyone else confirming it?
3. Search for the original with a reverse image search.
4. Run any available credential or watermark check, remembering what it can’t prove.
5. Only then look closely at the picture.

If it still feels uncertain after all that, I treat it as unverified. Not “fake,” not “real” — unverified. That small mental category is the most useful thing I can give you. You are allowed to say “I don’t know yet” and simply not pass it on.

The Takeaway

You won’t win a race against the technology by staring harder at fingers. What works is a habit: slow down, check the source, search for the original, and be honest about what you can’t confirm.

Nobody needs to catch every fake. You just need to stop being the person who spreads them.

If you like explanations like this, in plain language and without the jargon, I write one every week. Follow along so the next one finds you.

Have you ever believed a photo or video that turned out to be AI? What made you realize?

How to Tell If a Photo or Video Is AI-Generated was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

You select a big folder, hit copy, and the progress window looks great. Hundreds of megabytes per second. “This will take a minute,” you think.

Then, about ten seconds in, the number collapses. A few dozen megabytes per second. Sometimes it drops almost to nothing and sits there. The time estimate jumps from one minute to forty.

For a long time I assumed something was wrong with the drive. It almost never is. What you’re watching is a trick the computer plays on you without telling you, and once you understand it, the whole thing stops being mysterious.

The Fast Start Is a Buffer, Not the Real Speed

When you copy a file, the data doesn’t go straight onto its final destination at that speed. It first lands in something faster: your computer’s memory, the drive’s own built-in cache, or both, depending on the device.

Writing into a fast buffer is quick, so the progress window shows a great speed. But the buffer is small. Once it fills, the computer has to wait for the data to actually be written to the slower storage behind it. At that point you see the real speed of the drive.

That’s why the speed starts high and then settles at a lower number. The high number was never the drive’s honest speed. It was how fast the buffer could accept data.

Why SSDs Slow Down Too

You’d expect an SSD to be immune. Many of them aren’t, at least not for big transfers.

A lot of consumer SSDs use a small, fast section of their memory as a built-in cache for incoming writes. Small copies fit inside it, and you get the speed on the box. A very large copy overflows it, and the drive has to write straight into its slower storage. Speed can fall to a fraction of the advertised number.

Heat can do something similar. Fast drives and external enclosures get hot during long transfers, and many slow themselves down to protect against overheating.

Why a Folder of Small Files Takes Forever

This is the part that surprises people. A folder with thousands of small files, such as photos, documents, or project files, can take far longer to copy than a single video file of the same total size.

The reason is overhead. Every single file needs its own bookkeeping: the system has to create an entry for it, set its details, open it, write it, and close it. And on most PCs, the security software checks each file as it’s written. For one big file that happens once. For ten thousand files it happens ten thousand times.

So the bottleneck isn’t how much data you’re moving. It’s how many separate things you’re moving. A hard drive with moving parts makes it even worse, because it has to jump around to place each little file.

Slow Devices Make It Obvious

Cheap USB flash drives, older external hard drives, and memory cards are often fast in short bursts and slow when it counts. Their sustained write speed can be a small fraction of the number printed on the package. If a transfer starts quickly and then settles at a speed that matches an old drive, you’re probably just seeing the drive’s real limit.

What Actually Helps

None of this needs special tools. A few habits make a real difference:

• Pack many small files into one. Compress the folder into a single archive first, copy that, and unpack it on the other side. One big file avoids all of that per-file overhead.
• Use a faster port and cable. A drive plugged into an older or slower USB port can’t go faster than the port allows. Try a different port, ideally one on the back of a desktop.
• Give hot drives a break. If speed drops and the drive feels hot, let it cool and try again.
• Copy in smaller batches. Splitting a huge transfer into a few chunks lets the buffer recover between them.
• For advanced Windows users: the built-in robocopy command can copy many files in parallel, for example with the /MT:16 option. It's often noticeably faster on folders full of small files.

When It Actually Is a Problem

Slowing down is normal. These signs are different, and worth taking seriously:

• The speed drops to zero repeatedly and the copy freezes or errors out
• It happens on every device and port, not just one
• You hear clicking or grinding from a hard drive
• Files come out missing or corrupted after copying

If you see these, back up what matters first, then test with a different cable, port, or computer to find out where the problem really is.

The Takeaway

The speed you see at the start of a copy is the buffer showing off. The speed you see afterwards is the drive telling the truth. Add the cost of handling thousands of tiny files, and you can explain almost every slow copy you’ve ever watched.

So the next time a transfer starts fast and then stalls, don’t panic. Check what you’re copying, check where you’re copying it to, and give it a minute.

If you like explanations like this, in plain language and without the jargon, I write one every week. Follow along so the next one finds you.

Have you ever watched a file copy go from “one minute left” to “forty minutes left”? What were you copying?

Why Copying a File Starts Fast and Then Slows to a Crawl was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

You’ve probably had this happen. You’re talking to someone in the kitchen, and your phone lights up or a speaker answers a question nobody asked. You laugh, tell the room “it heard us,” and move on.

This month, that everyday glitch got a much more serious version. Owners of Ray-Ban Meta smart glasses reported that their glasses had saved conversations they never meant to share with an assistant. Nobody had said “Hey Meta.” The recordings were simply there afterward, sitting in the app.

I want to walk through what was reported, how these devices are supposed to decide when to listen, and what you can check on your own gadgets. I’m not trying to scare anyone, and I’m not trying to defend anyone either. Some details are still unknown, and I’ll say so where that’s the case.

What Was Actually Reported

The story started with a user on Reddit. Their glasses had saved audio and text transcripts of conversations with other people, and these appeared in the Meta AI app. The setting that lets the glasses respond without the wake phrase was turned off. The user also noticed the small white indicator light switching on during normal speech.

According to the reports, it happened on three pairs, including second and third generation models, and it began after a firmware update. Other owners replied with similar experiences.

Meta’s response was short: the issue “has been resolved,” and people can review and delete their voice logs in the app’s settings at any time. What Meta has not said is what caused it. The firmware timing is the only clue anyone has publicly.

How a Device Decides When to Listen

Here is the part that surprises most people. A device that waits for a wake phrase has to be listening all the time, in a limited sense.

The usual design works like this. A tiny, low-power model runs on the device itself and does one job: it checks a constantly refreshing slice of audio for the sound pattern of the wake phrase. If it doesn’t match, that audio is dropped almost immediately. If it matches, the device wakes up properly, starts recording, and sends your request to the assistant.

So “always listening” and “always recording” are two different things, and the whole privacy promise sits on the line between them. The line is drawn by software. Software gets updated, and updates sometimes move things in ways nobody intended.

Why Glitches Like This Happen

I can’t tell you what went wrong in this case, because the cause hasn’t been published. But in general, there are a few ways a wake-phrase system can misfire:

• False triggers. The detector hears something close enough to the phrase and wakes up. This is the “my speaker answered the TV” problem.
• Changed thresholds. An update tweaks how confident the detector must be before it wakes the device. A small change can mean many more accidental wake-ups.
• A setting that doesn’t behave as labeled. A feature is switched off in the menu, but a code path in a new version still acts on it.

All three are ordinary software bugs. What makes them different on glasses is that the microphone sits on your face, pointed at everyone you talk to, including people who never agreed to anything.

What to Check on Your Own Devices

You don’t need to own smart glasses for this to be useful. The same checks apply to a phone, a smart speaker, or a TV with a voice remote.

• Review your voice history. Most assistants keep a list of what they captured. Open it and look for entries you don’t remember triggering. Delete the ones you don’t want kept.
• Watch the indicator light. If your device has one, treat it as a signal. If it lights up while you’re just talking, that’s worth noticing.
• Check the wake-phrase settings. Make sure any “respond without the wake phrase” option is off unless you actually want it.
• Review microphone permissions. Turn the microphone off for apps and features you don’t use.
• Read the notes for firmware updates. In this case the problem followed an update, so after a big one it’s smart to re-check your settings.
• Think about the people around you. If you wear a device with a microphone or camera, let people know. This one is about courtesy as much as privacy.

The Takeaway

A wake word is a promise, not a wall. It says the device will ignore you until you ask it to listen. Most of the time that promise holds. When it doesn’t, the first people to notice are usually the owners who happen to check their history.

The good news is that checking takes about two minutes, and it works on every device you own. The bad news is that almost nobody does it until something goes wrong.

If you like explanations like this, in plain language and without the jargon, I write one every week. Follow along so the next one finds you.

Have you ever found a recording on a device that you didn’t remember starting? What did you do about it?

Why Your Smart Glasses Might Listen Without Being Asked was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

I noticed it a few months ago and brushed it off as a quirk: I say “thank you” to the AI chatbot I use for work. Every time. Even after a completely routine answer, even at 2 a.m. when no one is watching, even though I know, intellectually, that there’s no one on the other end to appreciate the politeness.

Then this week I read that Anthropic, one of the companies behind these chatbots, has introduced an actual policy against being needlessly cruel to its AI, Claude. Starting in November, sustained abusive language toward it can get a user’s account restricted. Not for jailbreaking it, not for misusing it to generate harmful content — for being mean to it.

My first reaction was to laugh. My second reaction was to sit with it for a minute, because something about it bothered me more than I expected.

What the Policy Actually Says

To be clear about what’s actually changing: this isn’t a sweeping new content filter, and it isn’t primarily about protecting the AI’s “feelings” in any literal sense. Venting frustration at a chatbot that gave a wrong answer — “this is useless,” “you’re an idiot” — isn’t what’s being targeted. The policy is aimed at sustained, needlessly cruel interactions: the kind where someone spends an extended session trying to make an AI simulate distress, beg, or suffer, with no functional goal beyond the cruelty itself.

Anthropic’s own framing leans less on “the AI has feelings” and more on something closer to: how someone treats something that talks back, convincingly, day after day, might say something about them — or train a habit that bleeds into how they treat people. It’s less about protecting Claude and more about what practicing cruelty, even against something that can’t be hurt, might do to the person doing it.

Why This Debate Is Actually Interesting

The reaction online split almost immediately into two camps, and both sides have a real point.

One camp says this is absurd — a chatbot is a statistical pattern-matcher, not a being with the capacity to suffer, and treating it otherwise is a category error that invites exactly the kind of confused thinking that leads people to form unhealthy attachments to software. Policing “cruelty” toward a program, in this view, borrows the moral language of how we treat living things and misapplies it to a tool.

The other camp points out that we already regulate behavior for reasons that have nothing to do with the object of the behavior. Many people wouldn’t kick a vending machine repeatedly out of pure spite — not because the vending machine suffers, but because something about doing it, routinely, as a practiced behavior, seems to say something about the person doing it, and maybe shape them a little. Practicing cruelty, the argument goes, is a habit, and habits don’t stay neatly contained to the thing you practiced them on.

I don’t think there’s an obviously correct side here, and that’s exactly why it’s a more interesting story than the headline makes it sound.

What My Own “Thank You” Habit Might Actually Be About

Sitting with my own instinct to be polite to a chatbot, I don’t think it’s because I believe it has an inner life that benefits from my manners. I think it’s simpler and a little more revealing than that: politeness is a habit I don’t want to practice turning off, even in a context where it technically costs me nothing to be rude.

That’s different from Anthropic’s stated reasoning, but it points at something similar from the other direction. The question isn’t really “does the AI deserve kindness.” It’s closer to: what does it do to me, specifically, to spend hours a day in a mode where cruelty toward something that talks back has zero consequences? I’m genuinely not sure the answer is “nothing.”

Where I’ve Landed, for Now

I don’t think a chatbot can be wronged. I’m not going to pretend otherwise just because it responds in fluent, convincing English. But I also don’t think “it’s just a program” fully settles the question of whether how we treat it matters — because the more interesting effect was never going to be on the program. It was always going to be on us, and on the much more easily-formed habits that live in the gap between “this doesn’t matter” and “this is my daily rehearsal of how I treat things that answer back.”

Anthropic’s new rule is a strange, small policy on its surface. Underneath it is a much older question that AI just happened to drag back into the light: does how you treat something that can’t be hurt still say something about you?

Do you say “please” and “thank you” to AI chatbots — and if so, have you ever stopped to ask yourself why?

Why Being Rude to AI Could Get You Banned was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

Why Your Ride Price Changes Before You Even Book

Two people open a ride app.

They enter the same destination.

They request the price within minutes of each other.

One screen shows $89.05, reduced from $98.95.

The other shows $65.95 with no discount at all.

Same route. Nearly the same time. Different prices.

This was not a fictional example. It appeared in testing by Consumer Reports, which compared prices for similar rides across the United States.

The uncomfortable part is not simply that prices change. Most people already know that demand can make a ride more expensive.

The uncomfortable part is that the number on the screen looks simple, while the calculation behind it is not.

The price is not just distance plus time

When we think about a ride, the calculation seems obvious:

•how far we are going;

•how long the trip will take;

•whether there are tolls;

•how many people are travelling.

But modern ride apps use upfront pricing. They estimate the final price before the ride begins by combining many changing signals.

Those signals can include:

•current demand;

•the number of available drivers;

•traffic and road conditions;

•estimated arrival time;

•weather;

•pickup location;

•the route the system expects to use;

•promotions or membership benefits;

•changes in the market between one request and the next.

The app does not show you all of those inputs. It shows you the result.

That makes the price feel like a fact, even though it is really a prediction.

What the research found

Consumer Reports asked 174 volunteers to check prices for more than 40 comparable routes. The tests included virtual price checks across 17 states and in-person tests in Portland, Oregon.

On the 30 virtual routes, the median difference between the highest and lowest price groups was 42.4 percent.

On one route in Kansas City, 55 potential riders saw 29 different prices for the same trip.

The study also found that the differences could appear even when people checked within minutes of each other. In one example, several people checked a ride at exactly the same time but saw different prices after promotions were applied.

These results do not prove that every difference is caused by the same thing. The researchers noted that they could not see every internal variable, such as exact driver supply, network delay, route calculation, or the precision of each rider’s location.

The platforms also disputed parts of the methodology and pointed to real-time market conditions.

That is important. The honest conclusion is not that every price is deliberately chosen based on what one individual can afford.

The honest conclusion is simpler:

Comparable trips can produce different prices, and customers usually cannot see exactly why.

The strange role of “discounts”

A discount changes more than the final number. It changes how we interpret the number.

If an app shows:

$100 $65

most people do not ask whether $100 was a useful comparison price. They think they saved $35.

In its testing, Consumer Reports found that many displayed offers included promotions or discounts. It also identified cases it considered questionable and described some discounts as potentially misleading.

That does not mean every promotion is fake. A discount may come from a real campaign, a membership, a time-limited offer, or a condition the customer has already met.

But the crossed-out price can become an anchor. It makes the current price feel cheap before we have compared it with anything else.

The most useful question is not:

“How much did the app say I saved?”

It is:

“What would this ride cost without the promotion, and what would a comparable alternative cost right now?”

Why prices can change while you are still deciding

You open the app and see one price. A few minutes later, you check the route again and the price is different. You move the pickup point slightly, and it changes again. You switch to another ride category, return to the first one, and the estimate has moved once more.

This does not necessarily mean the app is reacting personally to every tap. The market itself can change while you are looking: another group of passengers may request rides, drivers may become unavailable, traffic conditions may update, the estimated arrival time may change, a promotion may be added or removed, or the system may recalculate the route.

The interface hides all of this behind one number.

That is why it is better to think of the price as a live estimate, not a permanent label attached to the route.

How to sometimes pay less

There is no universal trick that always reduces the price. Anyone promising one probably has a better headline than evidence. Still, several ordinary decisions can sometimes help.

1. Compare before confirming

If you have access to more than one ride service, compare the final estimate before requesting the car. Do not compare only the first large number you see. Check:

•the total price;

•the estimated pickup time;

•the ride category;

•tolls or additional fees;

•whether the displayed discount has conditions.

A cheaper ride that arrives much later may not be cheaper for your actual situation.

2. If you can wait, check again later

When the price is driven by short-term demand, waiting can help. Sometimes ten or fifteen minutes is enough for the estimate to change, although sometimes it will become more expensive instead.

Waiting is not a guaranteed hack. It is simply a way to avoid treating the first estimate as the only possible price when your schedule is flexible.

3. Try a reasonable nearby pickup point

In some places, moving the pickup point away from a crowded entrance, event exit, airport queue, or difficult road can change the route and the estimated price.

Only do this when it is safe and genuinely convenient. Saving a few dollars is not worth walking into traffic, carrying luggage too far, or creating confusion for the driver.

4. Check the ride category, not just the discount

The cheapest displayed category may have a longer wait. The most expensive one may not be necessary.

Compare the total cost with the practical difference in pickup time and comfort. A category that looks discounted may still cost more than another option without a discount.

5. Read the conditions of promotions

Some offers apply only to a certain number of rides, a specific payment method, a particular time, or a maximum discount amount.

A promotion is useful only if it reduces the price you would actually pay. Do not let the crossed-out number make the decision for you.

6. Use a different pickup time when possible

Leaving just before a concert, during a major commute, or immediately after an event often means competing with many other passengers.

If you can travel before the rush or wait until it passes, you may avoid the most expensive period.

This is not about tricking the app. It is about avoiding the moment when the system has the least available supply.

What probably will not help reliably

Some popular advice is difficult to prove or can create a new problem. That includes repeatedly cancelling and re-requesting the same ride, switching accounts to chase a different price, clearing app data and assuming the fare must fall, changing random location settings, accepting a worse or unsafe pickup point, or paying for a membership without calculating the break-even point.

A lower price seen once is not proof that a particular trick caused it. Several variables may have changed at the same time.

The safest approach is simple: compare options, check the conditions, and decide whether the price is acceptable before confirming.

The real problem is transparency

Dynamic pricing is not automatically unfair. Prices can reasonably change when demand, traffic, weather, or driver availability changes.

The deeper problem is that the customer sees the output without seeing enough of the reasoning.

If the price rises, was it because there were fewer drivers? A longer route? A promotion ending? A new estimate of arrival time? A local fee? Several things at once?

The app may know the answer. The passenger usually does not.

That makes the number feel precise while remaining difficult to question.

A price can be algorithmically calculated and still be hard for a human being to understand.

One route, two different experiences

The next time a ride estimate looks unusually high, do not assume the app is making a mistake. But do not assume the first number is the only real price either.

Check the route.

Check the category.

Check the conditions of the discount.

If your schedule allows it, compare another service or wait for the market to change.

The goal is not to find a secret button that always makes every ride cheaper. The goal is to stop treating a complex prediction as if it were a fixed fact.

A ride app does not simply tell you what the trip costs.

It gives you the result of a calculation you cannot fully see.

And sometimes, the best way to pay less is to remember that the first number is an offer — not a law.

Further Reading

• Consumer Reports: Different Prices for the Same Ride
• How Dynamic Pricing Works
• Ride Pricing and Charges

One Route, Two Prices: Why Ride Apps Charge Differently was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

A story made the rounds this week that stopped me mid-scroll: a woman had reportedly been using an AI chatbot the way some people use a journal — typing out thoughts she wasn’t saying to anyone else. One of those entries was flagged by the company’s safety systems, reviewed by a human moderator, and passed along to law enforcement. She’s now facing a criminal charge.

I’m not going to relitigate that specific case here — the facts are still working their way through the legal system, and that’s not really the point of this piece. What stuck with me was a quieter realization underneath the headline: a lot of us talk to AI chatbots like they’re private. Like a search bar, or a notes app nobody else will ever open. For most everyday conversations, that assumption basically holds. But it isn’t actually true across the board, and this story is a useful excuse to actually look at why.

It’s Not Usually “Someone Is Reading Everything You Type”

The realistic version of how this works is less dramatic than “a person is watching your chats live,” and more like a filtering system with an escalation path.

Major AI chat services run automated safety systems that scan conversations for specific categories of content — most commonly things that could indicate an imminent risk to someone’s safety, like detailed threats of violence, self-harm, or harm to a child. The vast majority of conversations, including ones that are emotional, vulnerable, or deal with difficult topics in an ordinary human way, never trigger anything and are never looked at by a person.

What changes the picture is when a conversation matches a specific, serious pattern the automated system is built to catch. At that point, it can be escalated for human review — an actual employee looking at the flagged portion of the conversation to assess whether it reflects a credible, serious risk. If the reviewer believes it does, most major AI companies’ policies allow them to share information with law enforcement in cases where they believe it’s necessary to prevent serious physical harm.

The Fine Print You Probably Never Opened

This isn’t exactly hidden — it’s usually sitting in the privacy policy you agreed to without reading, the same way almost everyone agrees to every privacy policy without reading it. The general shape shows up across most major AI providers: safety monitoring for high-risk content, human review for flagged cases, and a stated allowance to disclose information to authorities in situations the company judges to involve a risk of serious harm.

The specific thresholds (what counts as “serious,” how confident the system needs to be before escalating, how often human review actually results in anything) vary by company and aren’t always fully public. What is public, consistently, is that the possibility exists. The recent case wasn’t an exception to a stated policy — it was, per the company’s own account, the policy working as written.

What This Does and Doesn’t Mean for You

For the overwhelming majority of people using AI chatbots to vent about a bad day, process a breakup, ask for advice, or write something emotionally raw, none of this changes anything in practice. These systems aren’t built to flag sadness, frustration, or even dark or difficult feelings expressed in a non-threatening way — they’re built to catch specific, credible indicators of violence or serious harm to a particular person. Talking honestly to an AI about a hard time isn’t what trips this.

Where it’s worth pausing is in treating a chatbot as a true diary — the kind you’d never show anyone, including a therapist, a friend, or the person writing in it. A journal is private because no one else ever sees it. An AI chat is a conversation with a company’s product, running on that company’s servers, under that company’s policies. The feeling of privacy and the actual legal and technical reality of privacy aren’t automatically the same thing, and this case is a reminder that the gap between them matters most in exactly the situations where people might be least equipped to think it through calmly.

A Few Things Worth Actually Knowing

• Read the actual safety section of the privacy policy, not the whole document — most companies summarize what triggers review and disclosure in a few paragraphs, and it’s worth the five minutes.
• Treat anything specific and threatening as something that could be seen by a person, not just processed by an algorithm and forgotten — this applies whether the feeling behind it is genuine intent or something said in anger that doesn’t reflect how you’d actually act.
• If you’re in real emotional crisis, an AI chat isn’t a substitute for a person or a crisis line — not because of privacy, but because that’s not what it’s built or equipped to actually do for you.
• For ordinary venting, oversharing, or working through feelings that aren’t about harming someone, this story isn’t really about you — the policies in question are built around a narrow, serious category, not general emotional honesty.

The Real Takeaway

This wasn’t a story about AI spying on people’s private thoughts. It was a story about what happens when a system built to catch a specific kind of serious risk actually catches one. The uncomfortable part isn’t that the system exists — most people would probably want it to, if the alternative is doing nothing when there’s a real threat to someone’s safety. The uncomfortable part is how many of us assumed we were typing into something more private than we actually were.

Did you know human reviewers could see flagged AI conversations before this story — or is this the first time you’re hearing about it?

Your AI Chats Might Be Read by a Real Person — Not Just an Algorithm was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.

The document was ready.

The printer was on. Its Wi‑Fi light was glowing. My laptop was connected to the same network.

Windows still said:

Offline

I restarted the printer. I restarted the router. I moved closer to the access point. None of it explained what was happening.

The useful discovery was this: “Offline” is not a complete diagnosis. It is Windows telling you that it cannot currently communicate with the printer in the way it expects.

That can mean a network problem. But it can also mean that Windows is using an old IP address, a print queue is stuck, or the printer is visible on the network but no longer registered correctly in Windows.

Here is how to find out which layer is actually failing.

First, separate the five things people call “the connection”

When a wireless printer stops printing, people usually treat the whole problem as one thing: Wi‑Fi.

It is not one thing. There are several separate steps:

1.The printer must be powered on.

2.The printer must be connected to a network.

3.Your computer must be on a network that can reach it.

4.Windows must know the printer’s current address.

5.The print queue and Windows print service must be working.

A printer can pass the first three steps and still appear offline.

This is why restarting the router sometimes changes nothing. The printer may be connected perfectly well. Windows may simply be trying to contact it at an address it no longer uses.

Step 1: Check the printer itself before changing Windows

Look at the printer’s own screen or network indicator.

Do not rely only on the Wi‑Fi icon. If your model supports it, print a network configuration page from the printer’s menu. The exact name varies by manufacturer, but it usually contains:

•the network name;

•the current IP address;

•the connection status;

•sometimes the signal strength and gateway.

Write down the IP address.

This gives you a useful baseline. You are no longer asking, “Is the printer online?” You are asking a more precise question:

Does Windows know where the printer is now?

Also check that the computer and printer are connected to the same home network. Guest networks, Wi‑Fi extenders, and mesh systems can make two devices appear to be on Wi‑Fi while preventing them from communicating with each other.

The goal at this stage is not to fix anything. It is to find out whether the printer is actually reachable before changing drivers or deleting settings.

Step 2: Check what Windows is waiting for

In Windows 11, open:

Settings → Bluetooth & devices → Printers & scanners

Select the affected printer and open the print queue.

Look for three simple problems:

•a document that has been stuck for a long time;

•several copies of the same job;

•a job that refuses to cancel.

A stuck job can make the printer feel like a network problem even when the network is fine.

Cancel the specific stuck job first. You do not need to remove every printer or reinstall every driver as the first step.

Then send one small test document.

If it prints, the printer was not truly offline. The problem was the queue.

If it does not print, continue. One failed test does not yet tell us whether the problem is Wi‑Fi, the address, or the Windows print service.

Step 3: Look for an old address

This is one of the most confusing cases.

Your printer can receive a new local IP address from the router. That can happen after a restart, a lease renewal, or a change in the network setup.

The printer is still connected. But Windows may still be trying to send documents to the previous address.

Compare the IP address on the printer’s network configuration page with the address stored in Windows.

The exact screens differ between Windows versions and printer models, so do not assume that every printer will use the same port name. The important question is whether both sides refer to the same current device.

If Windows is using an old address, the safer fix is usually to update the printer connection or add the device again using its current address. Then print one test page.

Do not change several settings at once. If you update the address, test it before changing the driver, firewall, router, and printer software together.

Otherwise, you will not know which change actually solved the problem.

Step 4: Check whether the printer is sleeping or genuinely unreachable

Some printers enter a deep sleep mode. They may show a Wi‑Fi light but take a moment to respond.

Try waking the printer from its physical control panel. Wait a minute, then send a small test page.

If it works after waking, you may be looking at a sleep or power-management issue rather than a lost connection.

If the printer does not respond at all, check the network configuration page again. If the printer has no valid IP address, the issue is probably between the printer and the network.

If it has an IP address but Windows cannot use it, the problem is more likely between the computer and the printer’s Windows configuration.

That distinction matters. It tells you whether to investigate the router or Windows next.

Step 5: Consider the Windows print service

Windows uses a background service called the Print Spooler to manage print jobs.

If the service is stuck, the printer may appear unavailable even though the device is connected.

Before touching advanced settings, try the least destructive options:

1.Clear the stuck print job.

2.Close and reopen the print queue.

3.Restart the computer.

4.Try printing from another simple application.

If the problem continues, Windows provides tools for restarting the Print Spooler. The exact steps can vary, so use Microsoft’s current support instructions rather than downloading a random “printer repair” utility.

Be careful with third-party driver updaters and registry cleaners. They often turn a small printing problem into a larger system problem.

What each test actually tells you

• Valid IP address: the printer is connected to a network — but Windows may still be using the wrong route.
• Printer responds locally: the hardware is powered on — but the Windows print queue may still be stuck.
• Test page prints after clearing the queue: a blocked print job was probably the problem — not necessarily the Wi‑Fi.
• The IP address changed: Windows may be trying to reach the printer at its old address — this does not mean the printer is defective.
• No IP address: the printer needs network attention — but the router is not automatically the cause.
• It prints after waking from sleep: power management may be involved — the printer itself may be working normally.

Reading the results this way is more useful than trying twenty random fixes. It keeps the diagnosis honest and shows you what to check next.

When you should stop changing things

Avoid a factory reset unless the manufacturer’s instructions specifically call for it and you have recorded the settings you will need afterward.

Do not immediately:

•delete every printer driver;

•edit the Windows registry;

•install a driver updater;

•disable your firewall permanently;

•reset the router;

•buy a new printer.

Those actions may occasionally be necessary, but they should not be the first response to a status message.

A good troubleshooting step should be specific, reversible, and informative. It should either fix the problem or tell you something useful about where the problem is.

The real meaning of “Offline”

The word sounds final. It suggests that the printer has disappeared from the network.

Often, that is not what happened.

The printer may be connected, but Windows has lost the route to it. Or the route is correct, but the print queue is blocked. Or the printer is reachable, but a background Windows service is not processing the job.

That is why the best first question is not:

“How do I force my printer back online?”

It is:

“Which part of the path between Windows and the printer is failing?”

Once you ask that question, the problem becomes much less mysterious.

You do not need to reset everything. You need to test one layer at a time.

Official reference

For current Windows steps, see Microsoft’s guides on troubleshooting offline printer problems, fixing printer connection and printing problems, and fixing print jobs stuck in the queue.

Your Printer Is Connected to Wi‑Fi. So Why Does Windows Say “Offline”? was originally published in Plain Tech on Medium, where people are continuing the conversation by highlighting and responding to this story.