The harmful effects of technology can be managed and mitigated. But not if we do not talk about them. Nor if we allow ourselves to be misled to focus on misleading narratives. In the chart The Elements of AI Ethics I map out harms that we are already seeing reports of, that have been ongoing and many of which were predicted before they happened. As a tool it can provide guidance and talking points for understanding how to prioritise your work with “smart” tools, and acknowledge that all teams that deploy or make use of AI need a mitigation strategy for many different types of harms.

This chart builds on The Elements of Digital Ethics (2021). The harms in the original diagram are all still relevant, and what this chart does is provide a focused overview of the types of harm we are seeing proliferate with the ongoing advancement of AI within different industries, and especially general-purpose and generative tools. You can refer to the original chart to get an idea of how a tool like this can be used in education and project work.

The following are brief summaries of each element to give you a better understanding of what they refer to. I am also putting together a workshop and talk around these elements. If that is something that would interest you and your organisation, let me know.

Download the chart in PDF format on the original post.

The Elements — Brief summaries

There are six overarching sections ( Organisation, Machine, Society, Human, Supervision and Environment), and at a total of 18 elements in the chart.

Organisation

These elements apply to the qualities and concerns of organisations that develop AI-enabled services and products in both the private and public sector.

Accountability projection

I am fond of the term “Moral outsourcing” as coined by Dr. Rumman Chowdhury. It refers to how makers of AI manage to defer moral decision-making to machines, as if they themselves have no say in the output.

By using the word accountability projection I want to emphasise that organisations have a tendency not only to evade moral responsibility but also to project the very real accountability that must accompany the making of products and services. Projection is borrowed from psychology and refers to how manufacturers, and those who implement AI, wish to be free from guilt and project their own weakness onto something else — in this case the tool itself.

The framing of AI appears to give manufacturers and product owners a “get-out-of-jail free card” by projecting blame onto the entity they have produced, as if they have no control over what they are making. Imagine buying a washing machine that shrinks all your clothes, and the manufacturer being able to evade any accountability by claiming that the washing machine has a “mind of its own”.

Machines aren’t unethical, but the makers of machines can act unethically.

Monoculture

Today, there are nearly 7,000 languages and dialects in the world. Only 7% are reflected in published online material. 98% of the internet’s web pages are published in just 12 languages, and more than half of them are in English. When sourcing the entire Internet, that is still a small part of humanity.

76% of the cyber population lives in Africa, Asia, the Middle East, Latin America and the Caribbean, most of the online content comes from elsewhere. Take Wikipedia, for example, where more than 80% of articles come from Europe and North America.

Now consider what content most AI tools are trained on.

Here is what i originally wrote about this element:

Through the lens of a small subset of human experience and circumstance it is difficult to envision and foresee the multitudes of perspectives and fates that one new creation may influence. The homogenity of those who have been provided the capacity to make and create in the digital space means that it is primarily their mirror-images who benefit — with little thought for the wellbeing of those not visible inside the reflection.

Deception by design

Many AI tools are consciously built to provide the illusion of talking with something that is thinking, considering or even being remorseful. These design decisions actually feed into the dangerous mindset of regarding the tools as sentient beings. As mentioned, this allows manufacturers to evade responsibility, but it also contributes to complex trust and relationship issues that humans have previously only had with actually sentient beings. We know very little about what it does to people’s emotional life and well-being in the long term to regularly interact with something that is designed to give the appearance of having human emotions when it in fact does not.

There are many reasons to assume that people can build strong bonds to tools that produce unpredictable output. And the number of people who need professional help to manage their mental health, and instead turn to AI chatbots, is of course growing fast. Sometimes they are even encouraged to.

• 'He Would Still Be Here': Man Dies by Suicide After Talking with AI Chatbot, Widow Says
• Eating Disorder Helpline Disables Chatbot for 'Harmful' Responses After Firing Human Staff

Power concentration

When power is with a few, their own needs and concerns will naturally be top of mind and prioritized. The more their needs are prioritized, the more power they gain. Three million AI engineers is 0.0004% of the world’s population.

Fearmongering sentience claims

There is an increasing number of reports coming out to describe a future, closer every day, wherein an AI takes over the world and makes humans obsolete and expendable. It is often fuelled on by the biggest names in the industry. By using exaggerated rumors of impending danger they manage to hold attention and control the narrative.

The claims have the effects of:

• Pulling attention away from the real harms that are happening
• Boosting trust in the validity of system output because they are framed as “all-powerful beings”
• Supporting the idea of accountability outsourcing (to the machines)
• Giving more power to the doomsayers because they are the only ones who can save us, and thus also the ones who can tell lawmakers how to regulate

Imagined Harm

In the chart I have chosen to make this last point overtly clear, by showing how the element “It will kill us all!” is an imagined harm with no evidence to back it up. It’s a harm in itself to push this narrative.

Machine

As more and more decision-making is automated, more humans must bow to algorithms, machine learning and obscurity. These elements describe the inherent qualities of machines that exacerbate potential harm.

Acceleration of bias and prejudice

One inherent property of AI is its ability to act as an accelerator of other harms. By being trained on large amounts of data (often unsupervised) — that inevitably contains biases, abandoned values and prejudiced commentary — these will be reproduced in any output. It is likely that this will even happen unnoticeably (especially when not actively monitored) since many biases are subtle and embedded in common language. And at other times it will happen quite clearly, with bots spewing toxic, misogynistic and racist content.

Invisible decision-making

The more complex the algorithms become, the harder they are to understand. As more people are involved, time passes, integrations with other systems are made and documentation is faulty, the further they deviate from human understanding. Many companies will hide proprietary code and evade scrutiny, sometimes themselves losing understanding of the full picture of how the code works. Decoding and understanding how decisions are made will be open to infinitely fewer beings.

And it doesn’t stop there. This also affects autonomy. By obscuring decision-making processes (how, when, why decisions are made, what options are available and what personal data is shared) it is increasingly difficult for individuals to make properly informed choices in their own best interest.

Society

The advent of digital has repercussions for society at large, influencing value-systems, opportunity, fear and safety to name a few. Things that were not previously possible are top of mind and often framed in savior-like qualities promised by digital transformation (and today most often by AI). New-found abilities and efforts to record and measure everything in numbers afford less attention to spirituality, compassion and autistic sensitivity.

Values detachment

Efficiency that involves handing over certain tasks to machines necessarily means that we give something up. We detach something. While it is often argued that we are giving up the mundande and tedious to live more happy and enriched lives this is not always the case. We need to keep an eye on what we are losing when we are gaining something else. By deferring even decision-making to machines (that were built by someone else and trained on an obscure mass of content) many people are incentivized to giving up doing work that may have been more important than the decisions themselves, e.g. speculation, reasoning and rumination. What we value changes and it is not obvious that we are handing over even the concept of our own value-based decision-making to an external party.

If societal values change over time, how will they change once they are part of the machine’s inner workings?

Acceleration of misinformation

The one harm that most everyone appears to be in agreement about is the abundance of misinformation that many of these tools will allow for at virtually no cost to bad actors. This is of course already happening. There is both the type of misinformation that troll farms now can happily generate by the masses and with perfect grammar in many languages, and also the type of misinformation that proliferates with the everyday use of these tools, and often unknowingly.

This second type of misinformation is generated and spread by the highest of educated professionals within the most valued realms of government and public services. The tools are just that good at appearing confident when expressing nonsense in believable, flourished and seductive language.

A lawyer used ChatGPT for legal filing. The chatbot cited nonexistent cases it just made up

One concern that does come up and is still looking for an answer is what will happen when the tools now begin to be “trained” with the texts that they themselves have generated.

The AI feedback loop: Researchers warn of 'model collapse' as AI trains on AI-generated content

Human

While most aspects of AI ethics are related to negative impact on human wellbeing, certain elements hit closer to home and are more direct and tangible in how they impact human wellbeing.

Acceleration of fraud & deepfakes

Tools that generate believable content with convincing wording, or tools that can imitate your voice or even your likeness, open up the playing field for fraudulent activities that can be harmful to mind, economy, reputation and relationships. People can be made to believe things about others that simply are not true, or even fall prey to the illusion that a family member is talking to them over the phone even when it’s a complete stranger.

• Found through Google, bought with Visa and Mastercard: Inside the deepfake porn economy
• 'Mom, these bad men have me': She believes scammers cloned her daughter's voice in a fake kidnapping | CNN

Acceleration of injustice

Because of systemic issues and the fact that bias becomes embedded in these tools, these biases will have dire consequences for people who are already disempowered. Scoring systems are often used by automated decision-making tools and these scores can for example affect job opportunities, welfare/housing eligibility and judicial outcomes.

• AI tools fail to reduce recruitment bias - study
• Ethics of AI-Enabled Recruiting and Selection: A Review and Research Agenda - Journal of Business Ethics

Content moderator trauma

In order for us to avoid seeing traumatitizing content when using many of these tools (such as physical violence, self-harm, child abuse, killings and torture) this content needs to be filtered out. In order for it to be filtered out, someone has to watch it. As it stands, the workers who perform this filtering are often exploited and suffer PTSD without adequate care for their wellbeing. Many of them have no idea what they are getting themselves into when they start the job.

https://time.com/6247678/openai-chatgpt-kenya-workers/?ref=axbom.com

Data / Privacy breaches

There are several ways personal data makes its way into the AI tools. First, since the tools are often trained on data available online and in an unsupervised manner, personal data will actually make its way into the workings of the tools themselves. Data may have been inadvertently published online or may have been published for a specific purpose, rarely one that supports feeding an AI system and its plethora of outputs. Second, personal data is actually entered into the tools by everyday users through negligent or inconsiderate use — data that at times is also stored and used by the tools. Third, when many data points from many different sources are linked together in one tool they can reveal details of an individuals’ life that any single piece of information could not.

11% of data employees paste into ChatGPT is confidential - Cyberhaven

Instigation of self-harm and violence

There have been several witness reports of AI chatbots encouraging or endorsing a human operator’s abuse towards themselves or others. Will manufacturers be able to provide a guarantee as to the complete avoidance of this in the future? And if not, what other mitigating factors need to be designed into the tool?

• Who is to blame: Can an AI Chatbot be convicted of inciting suicide?
• The Internet's New Favorite AI Proposes Torturing Iranians and Surveilling Mosques
• India's religious AI chatbots are speaking in the voice of god - and condoning violence

Supervision

This section relates to the difficult practice of overseeing, regulating and managing the potential negative impact related to AI.

Obscured data theft

To become as “good” as they are in their current form, many generative AI systems have been trained on vasts amount of data that were not intended for this purpose, and whose owners and makers were not asked for consent. The mere publishing of an article or image on the web does not imply it can be used by anyone for anything. Generative AI tools appear to be getting away with ignoring copyright.

At the same time there are many intricacies that law- and policymakers need to understand to do a good job of reasoning around infringements on rights or privileges. Images are for example not duplicated and stored, but rather used to feed a computational model, which is an argument often used by defenders to explain how generative AI is only “inspired” by the works of others and not “copying”. There are several ongoing legal cases challenging this view.

Sam Altman and the Sorcerer's Stone

Regulatory avoidance

AI can be seen as avoiding regulations by the sheer amount of oversteps being performed through their deployment and use. One possibility is that misuse of other people’s content may have to be normalised because oversight is unattainable, unless more constraints are quickly placed on “data theft”, deployment and use. The fact that so many of these companies can release tools based on the works of others, without having to disclose what they are based on, can only be seen as a moral dilemma of vast proportions. Moreover, what personal data is hidden to be abused or misused in future output is not only unknown to the operators of the tools, but to a large extent also unknown even to the makers of the tools themselves.

The fact that nations have different laws brings new obstacles to implementation across borders. OpenAI CEO Sam Altman for example expressed that he would pull out of the EU if he has to abide by proposed regulation. [update: Two days later he backpedalled those threats.]

Environment

The elements in this section refer to problems that are created and/or overlooked during the making of digital tools. They are often ignored because they are considered secondary conditions and do not relate directly to the intent of what is being created. Sometimes seen as untouchable or unavoidable, these elements still need to be addressed or managed to mitigate negative impact.

Supply chain neglect

To realise digital services and solutions we need both software and hardware. Behind the production of these there can be any number of oppressive relationships and exploitation of workers. One example is the need for cobalt, used in lithium batteries, which is often mined under unfair conditions in oppressive environments reminiscent of slavery. Failing to consider one’s own part in the supply chain required to deploy digital services is to ignore accountability for potential harm to others.

Anatomy of an AI System

Carbon costs

The energy required to source data, train the models, power the models and compute all the interactions with these tools is significant. While exact figures are often kept under wraps, there have been many studies into the large environmental cost of developing AI and measures that need to be taken. With this in mind, it is apt to ask if every challenge really is looking for an AI solution. Will consumers perhaps come to see the phrase “AI-powered System” in the same light as “Diesel-Powered SUV”.

The carbon impact of artificial intelligence - Nature Machine Intelligence

The hopeful part

The reason I want to contribute to the understanding of the types of harm entangled with the growth of AI is because all the harms are under human control. This means that we can talk about them and address them. We can work to evade or minimize these harms. And we need to demand transparency from manufacturers around each issue. Rather than pretending new technology is always benevolent because it will “improve lives” (definitely some lives more than others), surfacing these conversations is a good thing.

If you agree, I always appreciate when you can help me by sharing my posts and tools in your own networks. And I am always happy to hear about your own thoughts, perspectives and experiences. 😊

References and further reading can be found on the original post.

Originally published at https://axbom.com on June 2, 2023.

I have developed this communication model as a guide for conversations and strategic work aimed at improving AI tools. Note that the focus here is on human interaction with the tool. It is about the components that affect input data (from the human) and output data (from the computer), and by extension how a human chooses to act based on the output data.

The diagram contributes a communication science perspective to better understand when, where and how in the use and application of these AI tools that one can make an effort to improve for example

• comprehensibility,
• satisfaction and
• benefit.

It is useful in implementation planning, design decisions, communicative efforts and training.

What follows is a summary of the various components. I start on the left side of the model, with an operator who wants to accomplish a task. The word o perator is used here to describe a human who controls an interaction with a machine.

Although many examples are based on large language models, this communication model works for all systems where a human asks a question, or gives an instruction, to a computational model.

Download the communication model in PDF and PNG format on the original post.

The humanity of the operator

When a human is to interact with a computer, we need to keep in mind several aspects of the human’s circumstances:

• Expectations. How have the person’s expectations been formed and how well do they match what is possible in the given situation? How can we help keep these expectations at a good level?
• Ability. Given the person’s experience, knowledge and physical and cognitive capacity, how well equipped is the person to interact with the computer — for example, when it comes to phrasing inputs and prompts in a way that makes the computer produce a result they are satisfied with. How do we ensure that the tool is adapted to the operator’s ability?
• Context. What context is the person currently in and might this, for example, affect their ability to act as they themselves would like. How do we work with understanding context and taking this into account?

Input of question, or input data

When the person is asked to enter their question, request — or prompt — it is often done in a way that leaves room for interpretation.

• Intent. Does the person have an intent that we can subsequently check and verify. How do we help the operator phrase their intent?
• Code. Can the person encode their intention, linguistically, in a way that makes it, for example, distinct or useful to the computational model in a way that minimizes ambiguity. What support can help make the coding of the question more accurate?
• Expression. Does the input contain an expression that can affect the computational model in any direction? For example, this could pertain to formal, casual or childish language. How permissive or guiding should the tool be regarding these expressions?

The computational model starts its process

In the computational model itself — designed to accept the human input — many things are of course happening. Some of the most important aspects to keep in mind are:

• “Black Box”. It is hidden from the human how this works. Much is left to imagination and expectations built up by external circumstances. Even if the model used is released as open source, most people don’t have the ability to look at that code and understand what’s going on. You have to build your own story in your head to deal with the uncertainty. Which stories can be helpful and which stories mislead unnecessarily?
• Interpretation. There is a machine interpretation of the human phrasing, but not in the way that one human interprets another. When we are talking about a language model, some form of statistical analysis is generally used to simply calculate what are the most likely words and sentences that fit with what the person has entered. This is based on the data used to “train” the model. How confident are we with the interpretation for the purpose we want the tool to work?
• Computation (over truth). A calculation of the most probable sentences means, among other things, that even if the model has been “trained” on — and only been given access to — pure factual texts, it will find combinations of sentences that are probable but not necessarily true. There is no active check to increase the truth of the computation, and truth is not something that can be determined by the model on its own. The AI model is a computer program that follows instructions. Do we have a good grasp of what the computation actually does so that we do not promise things it doesn’t, and can not, do?

One way to describe this is that language models are often optimized so that the answers are perceived as true, but not that they are true. It is a language model, not a knowledge model. The manufacturers often prefer the answers to also be experienced as if they were written by a human… but we already know that is not true at all.

Consideration of the information store

The information upon which the computational model was “trained” is relevant. In the communication model, this is noted in the information store box, which also includes other information that can potentially contribute to the computational model.

When hearing stories about ChatGPT, it can sometimes sound like the tool has access to the entire Internet as a source of information for providing answers. But of course it’s nothing like that. It is not possible to compress all the content of the Internet in a database and let a computational model ask questions of the database. However, the tool may have been trained on large amounts of content found on the Internet.

The AI tool has been fed with these texts so that the computational model will receive more data to adjust how computations should be processed, and at best become more reliable, at least statistically. In the case of a language model, it is these calculations that select words so that the tool can imitate texts written by a human. Being “trained” with large amounts of information makes the computational model theoretically more adept at imitating human conversation and human writing.

The term Large Language Models (LLM) comes from the fact that the amount of information they were trained with is just this: very large. The idea is that the larger amount of information used, the “better” predictions the language model will make when it has to determine which answer, or output from the model, matches the question, or the input.

Similarly, we can see that tools that generate art can do so in a way that becomes more believable to humans the more human artworks the computational model is trained on.

An example of an area where language models are clearly useful is speech recognition. Some words or concepts are pronounced differently by different people, or can be drowned out by background noise. A language model can then help determine the probability that certain words are being said, based on all the others words being said, assisted by a computational model that has been trained on a lot of information. The purpose of the language model in this case is to completely copy what the human says and transform it into text. It becomes very clear what the person’s intent is, and it is possible to immediately verify whether the result matches both code and expectations.

It is different with language models where the goal is conversation, or where the goal is not very clearly defined. Then all sorts of problems can arise.

The following parameters do not always imply errors, but may be relevant depending on the purpose of a computational model. When you understand the enormous amount of information that a model is trained on, you also realize that the people who manufacture the model obviously do not have the opportunity to read the content themselves and somehow value its relevance. In the industry this is called unsupervised learning.

• Bias. The texts that are used may contain an enormous amount of prejudice of various kinds, including extreme opinions and incitement against ethnic groups. It is not only about overt racism but also more subtle inferences — that men deserve more pay than women, for example. In fact, given the sheer amount of content that AI tools sift through, these skewed views can be picked up by material that many people would at first glance consider neutral.
• Since a lot of material is needed for a model to become “better”, this always means that a great deal of historical material is used, where naturally there are biases. Societal values change over time but in old material they are locked in the era in which they were produced.
• Errors. Of course, clear factual errors are also found in the material that is used for training. But it’s important to remember that this is not primarily why language models can so often give wrong answers. Wrong answers are given because the models do not aim to generate facts, but aim to imitate human language. But just as a flat world map can contribute to a distorted understanding of what our planet looks like, inaccuracies expressed as truths for decades can contribute to these appearing in statistical calculations.
• For example, art forms originating in what is now Nigeria have long been attributed in European writing to Greek origins. It is in itself prejudices that lead to assumptions that lead to fallacies expressed as truth. Many such patterns, such as a historical reluctance to name African artists, can lead to a language model being unable to find statistically reliable ways to describe art and other culture of African origin. There are far more texts that identify Yoruban metal sculptures as Greek, although in recent years there has been a better understanding and the correct origin is given. So purely statistically — if you look at the amount of texts asserting the false perception — the sculptures are Greek.
• Constraints. We know that it is not the entire Internet that is used as an information source. Constraints have to be in place. Someone selects sources of information. Someone chooses a language. Someone chooses time periods for the source material, because it has to start and end somewhere. And all these constraints, of course, create their own biases because you can never claim that the source (especially in large models) is either representative or neutral.

Today there are more than 7,000 languages and dialects in the world. Only about 7% of these are reflected in online material. About 98% of web pages on the Internet are published in only 12 languages, and more than half of these are in English.

About 76% of internet users live in Africa, Asia, the Middle East, Latin America and the Caribbean — but the majority of content comes from elsewhere. On Wikipedia, for example, more than 80% of the content comes from Europe and North America.

Weaknesses in the information store can lead to the erasure of entire regions, cultures and languages from being represented in probable answers.

When using AI tools, one must pay attention to their weaknesses and how old values can be reinforced by them.

The response is presented

After the calculation is completed, the answer is presented to the operator. It is important to remember now that here too many important decisions are made by the manufacturer. There is not only one way to present answers, but of course an infinite number of ways. It is important to try and find out the manufacturer’s intent with the method that has been chosen.

• Computed response. Are we getting any clues as to why the output looks the way it does and how the query itself affected the output? For those who want to understand not only the answer, but also the reason for the answer, there is an opportunity here to give information about how well the answer fits the statistical model, or ask for more information to supplement and improve. It’s unfortunately rare that we see this form of transparency and support. Of course, this also reminds us that we are dealing with a machine, and if the manufacturer wants to hide that fact they also want to hide everything that reminds the operator of this. Humans and AI are often talked about as entities that should work together for the best results, and there are design decisions to be made here that can invite that collaboration. What parts of the computation are relevant for the operator to understand and know about?
• Code. Again, the answer, or output, must also be presented in a way that is easy to absorb but also matches the input and purpose of the tool. At best, we can verify that the output is reasonable and relevant based on the input. The weakness is that an incredibly well-formulated answer (human impersonation) can in itself contribute to the fact that the validity of the answer is not as easily questioned. If there are also no visual signals about how the answer is calculated, there is even less incentive to question it. How can the code be adapted to make the operator’s interpretation relevant and useful?
• Expression. In the same way that humans can write in a particular manner, so too can the output match it. Here, the manufacturer can choose to match the input, follow given instructions or find an entirely own tonality. Microsoft’s latest AI tool was criticized for adding a lot of emoji symbols to its responses. It is of course the manufacturer’s decision that the expression looked that way, and it provides a good example of how expression is a large part of how the response is received. What level of self-expression is appropriate given the tool’s scope of use?

Managing the response

The operator who now receives the answer must judge for themself its relevance and credibility.

• Decoding. Of course, being able to read the answer and understand all its components requires that the answer reveals all of this. Remember: the decoding is not only about being able to read and understand the text of a given answer, for example. It is also about forming an idea of how the answer provides value based on the question, or if it does not live up to expectations at all. There is a balance between simplifying and offering everything. What support can be offered when the operator decodes to better understand what is doubtful and what is definite in the output?
• Experience. What one feels and perceives, as an effect of how the answer looks and is presented, also plays a role in trust and satisfaction. When the decoding exceeds expectations in its expression, or leads to new thoughts and ideas, trust and satisfaction can be favored even when the answer does not have a direct correlation to the intent that existed when the question was asked. Do we understand the totality of the experience in a way that we also understand why the operator reacts and acts the way they do?
• Regulation. Given how satisfied or dissatisfied the operator is with the answer, there may be reasons to invite different ways to proceed which means staying with the tool but now regulating the completed interaction to steer towards a different answer. In the model we see adjust, expand and undo as examples of how the operator proceeds by regulating the whole, or parts, of the question. Some tools also encourage feedback by asking if you are satisfied or dissatisfied with the answer. This may imply a feedback mechanism that is used in some mode to influence the computational model. In what ways can the operator help themself, others or the tool achieve a higher degree of usefulness based on an answer already given?

Action and consequences

The action now involves taking the answer and using it for an application in some completely different context. We therefore have to think a step further, outside the communication model.

• How is the answer used to create benefits or value?
• How can we better contribute to a format, and integrations, that make the answer useful in the next stage of application?
• Who owns the answer?
• How can the answer help?
• How can the answer hurt?

The questions asked here will need to be worked out in research and interviews, but also with goal-oriented foresight. We must dare to ask the difficult questions and imagine both the desired and undesired future scenarios in order to be able to contribute to appropriate measures that, for example, minimize risks and make them visible when necessary.

To be able to do this in a good way, we must have a clear intent and intended use for AI-based tools in our operations.

If we invest in using and integrating AI tools in a business and no one takes responsibility for the effect on people, or for defining and evaluating the benefit, then we have given up before we have started. That’s when the strategy becomes to wait until something goes wrong and extinguish fires under duress, without preventive work or planned management.

This communication model is available as your support when you want to actively work to achieve well-thought-out, sustainable and compassionate solutions within the space of chatbots and AI-powered question-and-response tools.

Further reading available on the original post.

Originally published at https://axbom.com on May 31, 2023.

Applied ethics isn’t a checklist. It’s about putting in the time and effort to understand risks to wellbeing with the express intent of avoiding, mitigating and monitoring harm. It makes sense then to assume that the open call to pause AI development is a good thing. Well, yes — but no.

Let’s talk about some things that are going on with the touted open letter signed by the likes of Elon Musk, Steve Wozniak, Yuval Noah Harari, Andrew Yang, Gary Marcus and Tristan Harris. More than an open letter it is in many ways a letter of misdirection. And who authored it exactly?

Some points brought up in this story

• The host organisation
• Essentially no mention of all the current harm
• Boosting the idea of sentience
• Misleading citations
• The choice of “6 months”
• We vs. they

Read the full post:

Why that open letter urging an AI development pause is problematic

It’s never more easy to get scammed than during an ongoing hype. It’s March 2023 and we’re in the middle of one. Rarely have I seen so many people embrace a brand new experimental solution with so little questioning. Right now, it’s important to shake off any mass hypnosis and examine the contents of this new bottle of AI that many have started sipping, or have already started refueling their business computers with. Sometimes outside the knowledge of management.

This post is also available as a podcast episode. The original article is on axbom.com.

AI, a term that became an academic focus in 1956, has today mostly morphed into a marketing term for technology companies. The research field is still based on a theory that human intelligence can be described so precisely that a machine can be built that completely simulates this intelligence. But the word AI, when we read the paper today, usually describes different types of computational models that, when applied to large amounts of information, are intended to calculate and show a result that is the basis for various forms of predictions, decisions and recommendations.

Clearly weak points in these computational models then become, for example:

• how questions are asked of the computational model (you may need to have very specific wording to get the results you want),
• the information it relies on to make its calculation (often biased or insufficient),
• how the computational model actually does its calculation (we rarely get to know that because the companies regard it as their proprietary secret sauce, which is referred to as box), and
• how the result is presented to the operator* (increasingly as if the machine is a thinking being, or as if it can determine a correct answer from a wrong one).
• The operator is the one who uses, or runs, the tool.

What we call AI colloquially today is still very far from something that ‘thinks’ on its own. Even if texts that these tools generate can resemble texts written by humans, this isn’t stranger than the fact that the large amount of information that the computational model uses is written by humans. The tools are built to deliver answers that look like human answers, not to actually think like humans.

Or even deliver a correct answer.

It is exciting and titillating to talk about AI as self-determining. But it is also dangerous. Add to this the fact that much of what is marketed and sold as AI today is simply not AI. The term is extremely ambiguous and has a variety of definitions that have also changed over time. This means very favorable conditions for those who want to mislead.

Problems often arise when the philosophical basis of the academic approach is mixed with lofty promises of the technology’s excellence by commercial players. And the public, including journalists, of course cannot help but associate the technology with timeless stories about dead things suddenly coming to life.

It’s almost like that’s the exact association companies want people to make.

We love confident personalities even when they are wrong

Many tech companies seem so obsessed with the idea of ​​a thinking machine that they go out of their way to make their solutions appear thinking and feeling when they really aren’t.

With Microsoft’s chatbot for Bing, for example, someone decided that in its responses it should randomly shower its operator with emoji symbols. It is the organization’s design decision to make the machine more human, of course not something that the chatbot itself “thought of”. It is — no matter how boring it sounds and no matter how much you try to make it “human” by having it express personal well-wishes — still an inanimate object without sensations. Even when it is perceived as “speaking its mind”.

OpenAI’s ChatGPT, in turn, expresses most of its responses with a seemingly incurable assertiveness. Regardless of whether the answers are right or wrong. In its responses, the tool may create references to works that do not exist, attribute to people opinions they never expressed, or repeat offensive sentiments. If you happen to know that it is wrong and point this out, it begs forgiveness. As if the chatbot itself could be remorseful.

Then, in the very next second, it can deliver a completely new and equally incorrect answer.

One problem with the diligent, incorrect answers is of course that it is difficult to know that ChatGPT is wrong unless you already know the answer yourself.

Both the completely incorrect answers and the bot’s ability to completely change its mind when pointing out errors, are not strange in any way. This is how the computational model works. You’ve probably also heard this called a language model, or Large Language Model (LLM), but in practical terms it’s still numerical processing of observed data that determines what the tool prints out as a response.

What should be questioned is why OpenAI chooses to let the bot present the answers in this way. It is entirely according to the company’s design. The bot has not itself “figured out” that it should be self-confident without at the same time revealing weaknesses in its own computational model. That has been OpenAI’s decision. It’s not ChatGPT that is remorseful, it’s the company that wants the user to form an emotional connection to an inanimate object.

The illusion has largely succeeded. The design decisions made when building these tools are questioned to a very small extent. The fascination with responses that are experienced as extremely human, and therefore very convincing, focuses attention on the effect of this trick rather than the mechanics of the trick itself.

“Look over here!”, as the magician would say.

Of course people are seduced. A number of design decisions have been made to seduce. Just as with social media or e-commerce. The tools are built by people who are extremely knowledgeable in psychology, linguistics and behavioral economics. The way the answers are presented purely in terms of form and layout is far from a coincidence. The manufacturers are just as keen to influence the operator’s appreciation of the tool — despite dubious answers — as they are, of course, to try to improve the rules governing the computational model.

A lot of work and time is spent on presenting results to make them as engaging and captivating as possible. We know very little about what it does to people’s emotional life and well-being in the long term to regularly interact with something that gives the appearance of being alive when it is not.

The inadequate consumer protection

Misleading, designing and steering consumers towards specific behaviors and feelings is of course far from a new phenomenon. It is also a phenomenon that has given rise to laws and organizations intended to protect consumers and help them avoid adversity in the marketplace.

Read the rest of this story on axbom.com

Originally published at https://axbom.com on March 20, 2023.

They are reasonably quick to produce and they make a world of difference for curiosity, insight and impact. When I help organisation visualise their ideas, I call it being pictosmart.

1) Visuals, charts and infographics

I use a visual language to summarise complex subject areas, explain connections and guide towards improvement. These visualisations have a way of making their way into posters, reports, educational material, keynote presentations and strategic planning. They become an important part of sustainable learning and change in many organisations.

Some examples of flowcharts, mindmaps, systems diagrams and more

To see some examples of the types of diagrams I am talking about, you can follow these links. Each graphic can also be seen in the gallery below the links.

• Infographic: The many branches of the Fediverse
• Classification chart: The Elements of Digital Ethics
• Mindmap: Digital Ethics curriculum
• Systems map: The systemic complexity of managing a pandemic
• Flowchart: Manage impact in design
• Concept Sketch: Read the rest of this story on axbom.com

💡

Don’t know what Mastodon even is? Start over here.

The intent of this short list is to:

• Help steer you away from the biggest instances. Not only are the biggest instances contradictory to the idea of decentralisation, they also amass many users in short periods of time which can make them slow and unresponsive.
• Get you started. You will understand types of differences between instances, and perhaps after this feel more confident diving into the bigger directories of instances. Or you choose one from this list list and dive right in, which perhaps is the most valuable learning experience you can have. As a reminder: you can always move to another instance later, keeping your followers and follows. Your posts will however not carry over.

11 Mastodon instances to consider

⚠️

Everyone on all instances can follow everyone else on all other instances. Unless there are blocks and suspensions preventing those connections. Those blocks most …

Read the rest of this story on axbom.com

So while the underlying tech has changed many times in this quarter-century, there has been no new address to learn. No redirect needed. Readers could always find my writing on axbom.com (and eventually also axbom.se when I split into two different language versions).

The term wasn’t coined yet but the domain name was my key to the independent web, or indieweb.

💡

Indieweb refers to owning your own domain name and your own assets online. The idea is to avoid being dependent on commercial actors who may misuse your data, exploit it for harmful impact, or simply make it disappear. Read more on Grow your own services.

Read the rest of this story on axbom.com

I am well aware this still doesn’t cover all available tools, and ActivityPub isn’t the only protocol. The image is a simplification, as most models are. Given that new tools can be made available regularly it would be an insurmountable task to pursue a complete mapping. Hopefully this visual is enough to illustrate the amazing variety of services within the Fediverse, and will spark curiosity to explore further.

You can download the diagram in PNG format or as a PDF.

The many branches of the Fediverse (PNG)

fediverse-branches-axbom-PA01.png

963 KB

download-circle

ℹ️

This is for people who now have Mastodon accounts and have poked around a bit, but are still a bit confused about the whole experience.

1) Good apps

Metatext or Toot! on iOS and Tusky on Android are oft-recommended apps. One progressive web app (PWA) I’ve also enjoyed is Pinafore.

Remember it’s possible to just use the web interface. You can save the web address of your home instance as a shortcut on your phone. I myself prefer this because it maintains all the features that some apps forget to implement (like setting the language of your toot). It also helps me avoid notifications on my phone.

No account yet? For creating an account, start in a web browser first, before downloading an app. Also, beware of some apps renaming features to something that is unique for that app.

Yes there are “official” apps by the lead Mastodon developers but I rarely see them recommended.

2) Understand the concept of “instance”

You can think of instance as the web address at the end of someone’s user name. This is where they signed up and this is where their version of Mastodon is hosted. Examples of instances: mastodon.lol, floss.social, qoto.org,…

Read the rest of this story on axbom.com

1. Caryn

In 2017 a woman by the name of Caryn told this story. She had a friend who died. And she did not know, because of algorithms. Let’s call him Adam. Adam had fallen terminally ill and passed away within a timeframe of a few months. A mutual friend reached out to Caryn to ask her if she was aware that Adam had died. She was not.

An avid Facebook user, Caryn visited Adam’s Facebook page. They had met in-person once, but mostly interacted online for more than 15 years. She found a post he had made a month earlier, explaining his situation as he was in hospital. Caryn and several other people she reached out to had never seen this post.

Caryn understood what had happened. Because Adam did not post much, he mostly just liked and commented, the algorithm did not value his post enough. It chose to not show it to a number of his friends.

But she also realised what this could have meant for Adam. He may have passed away thinking that none of his online friends cared.

Read the rest of this article on axbom.com