Who decides what artificial intelligence should become? A laboratory may build a model, a company may place it inside a product, and millions of people may use it before anyone understands all its effects. In the Center for Humane Technology’s widely watched presentation The A.I. Dilemma, Tristan Harris and Aza Raskin argue that the way we develop and deploy AI deserves as much attention as the technology itself. Their warning is not a prediction carved in stone. It is an invitation to examine the incentives that steer our choices.
The talk was delivered in March 2023, during a period when generative AI suddenly entered ordinary life. Some demonstrations and forecasts in it belong to that moment. Its central question remains useful: when a system can shape information, relationships, work and public decisions, what would responsible development look like? We can take the concern seriously while testing each claim against evidence and acknowledging that AI also has real benefits.
Why a race changes the choices people make
Imagine several companies developing similar tools. Each fears that waiting will allow a rival to win customers, investors or attention. Even leaders who care about safety may feel pressure to release first and repair later. This is the race dynamic the presenters emphasize. It does not require any one person to be careless. The outcome can emerge from a system that rewards speed more visibly than restraint.
That pattern is familiar from earlier digital platforms. A recommendation system optimized for engagement can succeed at its narrow target while creating problems the target never measured. AI makes the stakes broader because one underlying model can be placed inside search, education, customer service, software development and personal conversations. The question is no longer only whether a model performs well in a demonstration. It is what happens after the demonstration becomes infrastructure.
The Center for Humane Technology describes the talk as a call to understand how competitive pressure and business models influence AI design. That is the organization’s perspective, not a neutral measurement of every AI deployment. It is strongest when treated as a framework for asking specific questions: What is being optimized? Who bears the cost when a system fails? Who gets to challenge a decision?
Capabilities are not the same as trust
A model can write elegant prose and still invent a citation. It can summarize a document and miss the one sentence that changes its meaning. It can appear conversational without understanding a person’s circumstances. Fluency encourages us to trust an answer before we have checked it. The more useful the tool becomes, the more important it is to notice that difference.
Trust should be earned through evidence matched to a purpose. A brainstorming assistant needs one level of reliability. A system used to guide medical decisions, assess applicants or control critical equipment needs a much higher one. The US National Institute of Standards and Technology’s AI Risk Management Framework asks organizations to consider validity, safety, security, privacy, fairness, transparency and accountability across an AI system’s life cycle. Those words become meaningful only when they lead to tests, monitoring and ways to respond when harm occurs.
This also explains why a single accuracy score tells so little. A system may work well for common cases while failing for a small group. It may be reliable in a laboratory and brittle after users adapt to it. It may give a correct answer while disclosing information it should protect. Good governance asks how a tool behaves in the setting where people actually rely on it.
Look for the people behind every automated decision
AI is often described as though it acts on its own. In practice, people choose training data, product goals, access rules, evaluation methods and the places where a system will be used. People also decide whether a human can review an outcome. Calling a decision “automated” must not erase the responsibility of those who designed and deployed the process.
Consider a school adopting an AI tutor. The useful question is not simply whether students enjoy it. Teachers need to know what it gets wrong, how it handles uncertainty, whose material it draws on and whether students can still practice independent thought. Parents and students need a clear account of what information is collected. The same reasoning applies to tools in workplaces and public services. A benefit in one area does not cancel a harm in another.
UNESCO’s Recommendation on the Ethics of Artificial Intelligence places human rights, dignity and oversight within a broader approach to policy. It does not hand us a universal yes or no answer for every tool. It does suggest that affected people should have a voice in decisions that shape their lives. That principle turns abstract talk about “human centered AI” into questions about consent, recourse and public accountability.
Neither panic nor passive optimism helps
The talk uses a sense of urgency. Urgency can be valuable when it moves people to investigate and improve systems. It becomes less useful when every risk is presented as equally likely or every future as inevitable. We should distinguish harms that can already be observed from plausible future risks and from speculation. Each deserves attention, but each calls for different evidence and different action.
Passive optimism has a similar problem. Saying that innovation will solve its own problems does not identify who will test a system, who will pay for safeguards, or how a person can contest a harmful outcome. Progress does not arrive simply because a product is new. It is built through decisions that can be inspected and changed.
A practical approach begins with the use case. Before adopting a tool, write down the problem it is supposed to solve. Compare it with a simpler alternative. Test it with people who understand the setting, including those likely to be affected by errors. Decide in advance when a human must intervene, what data can be retained, and what would trigger a pause or withdrawal. Then keep measuring after launch. These steps are less dramatic than a forecast about the end of the world, but they are where public trust is earned.
What can an ordinary reader do?
Start by asking for the source behind an impressive AI claim. Was the result reproduced outside the company that announced it? What task was tested, and what was left out? When you use an AI tool yourself, treat its answer as a draft to check, particularly when the stakes are high. Protect private information, look for original documents and keep your own judgment active.
You can also ask institutions for clarity. If a workplace, school or service uses AI to shape decisions, it is reasonable to ask what role the system plays and how a person can request review. A useful policy should be explainable in ordinary language. The aim is not to make every citizen a machine learning engineer. It is to keep decisions that affect people open to questions and correction.
The future is still a choice
The A.I. Dilemma is compelling because it connects technical power to human incentives. Its enduring lesson is that we should evaluate the system around a model as carefully as the model itself. A remarkable demonstration can inspire curiosity. It cannot substitute for evidence about how a tool behaves in daily life or who is accountable when it fails.
Watch the presentation with both openness and scrutiny. Notice where the speakers identify a concrete mechanism, where they offer a forecast, and where new evidence may have changed the picture since 2023. The future of AI is neither fixed by the first company to release a product nor guaranteed by hopeful language. It will be shaped by the questions people insist on asking and the standards they are willing to enforce.
Featured video: The A.I. Dilemma — Center for Humane Technology.
