How Different AI Horizons Shape Psychology’s Research, Policy, and Practice

People who notice AI progress now are noticing abilities in free versions of apps that were only present one or two years ago in the most advanced models, or weren’t present at all. If they are exposed to more rapid advancements through social media, paid versions, or work, they begin to see capabilities that look as good as or better than humans on some tasks.

I think current AI capabilities are progressing closer to what people consider artificial general intelligence: AI that is about as good as humans across most digital cognitive work. That makes it easier to recognize AGI as a possibility, or even an imminent reality. From there, the question becomes what happens when AI is smarter than humans by a large margin.

Those different AI horizons matter because they change the research, policy, and practical decisions psychology should be considering. We’re compressing the time between them.

What Happened This Week

On September 8, Jacob Coxon announced his resignation from Anthropic after doing pretraining research at both Anthropic and OpenAI. He accused the companies of racing irresponsibly toward self-improving superintelligence: “The people building AI earnestly believe that it could kill us all by the end of the decade.”

By September 9, Axios reported more than 115 million views on his resignation post. Interviews followed, including with WIRED. Anthropic’s alignment science lead, Evan Hubinger, publicly backed the concern and estimated the chance of human extinction from AI above 10% within the next decade.

Ted Lieu cited Coxon’s post in calling for passage of his bipartisan AI Kill Switch Bill. Bernie Sanders supported the warning. Sanders and Greg Casar had already announced plans on September 3 to ban superintelligence and temporarily pause advanced AI development, with a new federal agency and international cooperation.

Zvi Mowshowitz’s “Three AI Pills” provides useful shorthand for discussing and modeling these different views. Pill 1 means taking current AI capabilities seriously. Pill 2 means taking AGI seriously. Pill 3 means taking artificial superintelligence seriously: AI that can do approximately all the things better than humans within our lifetimes.

I think most people still haven’t really taken current AI capabilities seriously. But someone doing so now may effectively be closer to Pill 1.5. They are much more likely to recognize human-level AI because of what is already in front of them.

But a person can make a statement about superintelligent AI while still holding beliefs closer to current or human-level AI. Someone can fear AI taking over a power grid or helping humans create a bioweapon while still expecting it to remain jagged, dependent on humans, and limited in important ways.

AI is already jagged and superhuman in many ways. Someone might see spikes beyond human-level AI over the jagged edge without adopting the full superintelligence view. In the shorthand, I think of that as something like Pill 2.1.

Sanders’s proposed policy goes much further than conventional chatbot safeguards. But an extreme policy does not tell us someone’s entire underlying model. A person can support stopping superintelligence without having worked through AI research automation, accelerating science and engineering, robotics, and economic feedback loops.

I think people making statements about superintelligent AI while otherwise reasoning from current or human-level AI may not remain committed to the same response. Regulation may settle around testing, audits, liability, cybersecurity requirements, and government evaluation.

Except we’re compressing the time. Another capability event may arrive before the previous alarm has faded. People might repeatedly move toward the next horizon before they have time to move backward.

What This Means for Psychology

This belongs in every field, because everyone will be impacted, but doubly so in psychology. Psychologists have the ability to help guide AI development because of our training, research orientation, expertise across different areas, and multidisciplinary collaboration. APA has made that case too, describing roles for psychologists throughout AI development, from framing research questions to evaluating real-world impact.

Just as these horizons are compressing more broadly, that could also be true for psychologists. They see what AI is currently capable of through their own use and through studies. They can more quickly recognize AGI as a possibility and begin considering the risks and implications that follow.

But much of the research and professional discussion I’m encountering is about how people interact with ChatGPT, ethical note-taking, ethical consultation, and shared learning with AI. APA’s professional guidance addresses clinical decision-making, documentation, and patient engagement; it also has guidance specifically for evaluating AI scribes.

Basic versions of these uses were possible two or three years ago. There needs to be empirical research on their safety for patients and on psychologists’ own use. My concern is that we keep researching those capabilities while what is currently possible moves ahead. In the work I’m seeing, that creates a lag of roughly two or three years, which I don’t think is acceptable at this pace.

What are current capabilities? Increasingly capable coding agents, sustained work across multiple steps, and systems that use tools and maintain information across longer workflows. Those are already development priorities and available forms of agent support.

Getting back to my patients, consider a chatbot between sessions that has the session history, can answer questions, and can flag something for the therapist. Privacy, consent, reliability, and escalation would need to be addressed. But those are questions about how to build and evaluate something that is technically plausible now.

For case conceptualization, consider taking multiple inputs and developing several possible formulations, referencing the DSM and other documents, and retaining small pieces of information that matter in an unusual case. Earlier systems were terrible at parts of this. Now there is much more to evaluate: accuracy, cultural responsiveness, longer-term coherence, and whether the system actually uses the relevant details.

Then there is creating a program, researching a therapy, drafting a treatment manual, making visual handouts, taking meeting notes, and preparing presentations. An agent can potentially carry work across those steps, including related emails and other actions when authorized. How does that change psychologists’ work, their trust in the system, and what they need to check?

There are also live avatars. HeyGen, for example, offers interactive avatars with responses guided by a knowledge base or connected model. It can be a surreal interaction. It doesn’t have to pass a Turing test for us to start seeing something that looks more like a person talking to us over video than a conventional chatbot.

These are current-capability questions. We don’t have to wait for AGI to study them.

Psychology With Human-Level AI

If AI becomes about as good as humans across most digital tasks, it could potentially function as something close to a drop-in therapist over Zoom. We would still need evidence of therapeutic effectiveness, but the research and policy preparation would look very different.

That does not mean human psychologists disappear. Regulation, patient preferences, relationships, and trust would still matter. A lot of people will probably care that they are talking to a human and choose that. Other people might not care, or might prefer AI.

How do those multiple worlds coexist? And how does the field coexist with them?

There are questions about collaborative treatment, licensing and regulation, and disclosure of whether someone is interacting with a human or AI. What ethical responsibility do psychologists have to consult AI that is better than them at a relevant part of the work? What happens to standards of care when AI judgment becomes better than human judgment?

Those are questions that follow from taking human-level AI seriously.

Psychology With Superhuman AI

Now consider AIs that are smarter than humans by a large margin. That brings questions about persuasion, surveillance, existential risk, disempowerment, embodiment, and human enhancement.

A lot of people would say this is science fiction. But there are significant psychological implications.

Consider AIs that are better than psychologists at essentially every cognitive aspect of their jobs, except where people specifically prefer a human. How would psychologists oversee a system that understands and influences people better than they can?

What ethical choices would we have about using its persuasion techniques in treatment? Whose goals would those techniques serve? How could psychologists help evaluate whether AI is actually following what we ask it to do?

My own estimate of the existential risk from AI is about 10%. That relates to Coxon, but it relates to psychologists as well. Thinking about what alignment means is partly thinking about the prediction of behavior, and this is what psychologists do. Are we able to help predict the behavior of AI, maybe not only from a technical perspective, but also from a psychological perspective?

What would that look like? Should we think about AI control, Anthropic’s Constitutional AI, or other models? What values should be chosen, and who chooses them? Is this even possible? Are there examples from history of what this looks like? These are some of the biggest questions raised by superhuman AI once you take the actual AI safety risk seriously, and I think psychologists can be part of that research as well.

Then there are alignment questions about human values globally. How do democratic values interact with other values? Should we be imbuing an AI with these values, and who gets to decide?

There are also questions about AI itself. How much should we anthropomorphize a system that displays preferences, distress, identity, relationships, or the functional characteristics people associate with consciousness? Those displays would not establish that it has subjective experience. But what moral rights or ethical responsibilities might need to be considered?

Could AIs receive something analogous to psychological treatment, even if they were not considered conscious? Would it help them continue learning or function as better agents? Would they be so much smarter that only other AIs could counsel or modify them at a useful speed and scale? How would psychologists research that, and would some interventions be unethical?

These are the kinds of questions I mean by taking artificial superintelligence seriously.

As a psychologist in training, I want to stay grounded. I don’t know whether the transition from human-level to substantially superhuman AI takes a year, five years, or ten years. I expect it could be much faster than our institutions are preparing for.

I want psychologists to ultimately take artificial superintelligence seriously and understand those stakes. For this article, the point is that there are practical, policy, and institutional differences depending on our beliefs about current AI, AGI, and artificial superintelligence.

We can continue researching patient safety, privacy, bias, and psychologists’ own use while also developing a conditional research and policy agenda for those possibilities. We should not have to wait until each capability is already deployed before deciding what questions matter.

Process note: AI was used as a co-writing, research, editorial, and publishing collaborator. Estimated authorship: wording, including lightly cleaned speech, is 65–75% mine and 25–35% AI; ideas and substantive arguments are 90–95% mine and 5–10% AI; structure and order are 70–80% mine and 20–30% AI. These are rough editorial estimates, not measured text comparisons.