Skip to content
ChatGPT · Artificial intelligence · OpenAI o3Episode 063 · 22 June 2025 · 42:33

AI Promises a Therapist and an Honest Startup—and Receives Trust That Should Never Be Given Automatically

Central question

Why does AI receive the trust given to a therapist or adviser before the product can justify that trust?

What you take away

Draw the line between a useful AI conversation and the role of a therapist or adviser by considering the cost of error, verification, and the trust granted to the product.

Main threads

What to watch for

1Compare “AI Promises a Therapist and an Honest Startup—and Receives Trust That Should Never Be Given Automatically” with “AI therapist: hands-on experience”: they provide different criteria for judging the same issue.
2Test the conclusion from “Lawyers and fake AI citations: the London penalties - part 2/2” in your own use case—what actually changes in the process and what remains a promise.
3Before choosing a product or approach, record the constraint identified in “Scandal with Builder.ai: $1.5 billion start-up of Microsoft and SoftBank”.
4Define the owner of the outcome and the quality metric for the situation described in “ChatGPT O3 PRO”.
Signals to track afterwards
Watch for actions by Stanford and Ilnar Shafigullin that confirm or challenge the episode’s central claims.
Compare new launches and policy changes with “Lawyers and fake AI citations: the London penalties - part 2/2”: have access, quality, price, or constraints changed?
Check whether the scenario in “ChatGPT O3 PRO” becomes repeatable practice rather than a one-off demonstration.
Most useful for
Executives and managersEntrepreneursInvestorsCompany leadersUsers of healthcare servicesHealthcare teams

Key takeaways

00:00An AI therapist is convenient for an obvious reason: it is available at night, does not interrupt, and does not judge

The working conclusion from “An AI therapist is convenient for an obvious reason: it is available at night” is that it can be real support when no specialist is nearby, yet the model cannot see the patient's condition, carries no professional responsibility, and may confidently continue a mistaken line — an assistant and a therapist are not the same thing.

02:25The same mechanism of trust operates in business

The “AI therapist: hands-on experience” topic becomes clearer once this point is included: turning to a model to extend something you already understand and can verify is one thing; blindly trusting a confident answer where you cannot check it is another — all the more so because free apps often run on old, limited, hallucination-prone models.

10:25What changes in real work: the Stanford study on chatbot therapists

The working conclusion from “The Stanford study on chatbot therapists” is that Stanford researchers examined five popular chatbots marketed as cheap mental-health help and found they reinforce stigma (more toward alcoholism and schizophrenia than depression), can suggest dangerous actions in response to suicidal thoughts, and fail to recognize critical situations that require a human.

13:48Why context matters more than one metric: the DeepSeek scandal — Italy's antitrust case

The working conclusion from “The DeepSeek scandal: Italy's antitrust case” is that Italy's competition authority opened a case against DeepSeek for not warning users about the risk of inaccurate information — the first precedent of its kind, and it hits the shared problem: an answer looks more reliable than it is.

15:40How the issue moves from news to product: AI and vibe-coding — the risks of irresponsible programming

The “AI and vibe-coding: risks of irresponsible programming” scene leads to a working conclusion: when everyone suddenly “becomes a programmer” and generates apps in chatbots, responsibility for working code still rests with the person — you should be able to read what the model produced and understand what that code will do.

18:52Education often reacts in the opposite way by trying to ban AI from exams and assignments

The decision in “Lawyers and fake AI citations: the London penalties” depends on one criterion: London's High Court will penalize lawyers for fake AI-generated citations — in large research outputs a model confidently cites sources that do not exist, and responsibility falls on whoever signed and filed the document.

20:55Where the promise meets reality: will jobs vanish because of AI? The PwC study

The working conclusion from “Will jobs vanish because of AI? The PwC study” is that PwC's 2025 jobs barometer, drawn from hundreds of millions of vacancies, found that in the industries adopting AI fastest, jobs grow rather than shrink, wages rise twice as fast, and workers without a degree take on more complex tasks.

30:00In each case, the model amplifies a weakness that already existed in the system

The “Scandal with Builder.ai: $1.5 billion start-up of Microsoft and SoftBank” issue should be assessed with one constraint in mind: a persuasive founder, a polished deck, and talk of AI can create the image of a tech company even where the product and revenue do not support the valuation — the market easily starts buying not a system but a well-packaged version of the future.

34:48The new skill is therefore not merely knowing how to use ChatGPT

The boundary of the “ChatGPT O3 PRO” case is defined by this point: for the first weeks o3 Pro answered very slowly, then Deep Research sped up to a few minutes; but users notice o3 has gotten worse — the familiar cycle where a model launches strong and then quality slips — and on genuinely novel hard problems the models still fail, remaining a tool rather than a replacement for a specialist.

What this episode is about

People already tell models things they are unwilling to tell friends, investors believe persuasive presentations, and schools try simply to ban AI. In all three cases, the problem is the same: a polished answer or an attractive story looks more reliable than it really is.

An AI therapist is convenient for an obvious reason: it is available at night, does not interrupt, and does not judge. That can be genuine support, especially when no specialist is nearby. But the model cannot observe a patient’s condition, carries no professional responsibility, and can confidently continue along a mistaken path. An assistant and a therapist are not the same thing.

The same mechanism of trust operates in business. A persuasive founder, a polished presentation, and language about artificial intelligence can create the image of a technology company even where the product and revenue do not support the valuation. The story of a startup valued at one and a half billion dollars shows how easily investors and the market begin buying not a system, but a well-packaged version of the future.

Education often reacts in the opposite way by trying to ban AI from exams and assignments. A simple ban does not restore the old world. A student can use a model covertly, a teacher may not understand where assistance was used, and assessment becomes a contest among detectors. It is more important to redesign the task itself and require an explanation of process, sources, and decisions.

In each case, the model amplifies a weakness that already existed in the system. In psychology, it is a shortage of accessible help; in venture capital, reliance on narrative and trust; in education, evaluating an outcome without understanding the path. AI does not create the problem from nothing, but it sharply increases its scale and speed.

The new skill is therefore not merely knowing how to use ChatGPT. It is understanding the boundary of authority: where an answer can be accepted as a draft, where a professional is required, where documents are needed, and where a person must show their reasoning. The more convincing AI becomes, the less sensible it is to judge the system only by how naturally it speaks.

The more convincing AI turns into, the less sensible it is to judge the system only by how naturally it speaks.

Episode transcript

The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 110 segments: 49 identified, 1 mixed, 32 probable, and 28 unresolved.

Loading…