The Thing That Made Me Uncomfortable

Here is the part of the study I had to read slowly.

The drop that should give us pause

At the start of the study, about three out of every ten students in each group said they wanted to see a therapist for help.

Twelve weeks later, in the AI group, that number had gone down. Fewer students felt they needed a therapist (Shoshani et al., 2026).

In the no-help group, the number went up. They had not gotten better, so more of them wanted help.

In the human therapy group, the number stayed about the same.

So far, this might sound fine. People got better. They felt they needed less help. That is the good story.

But there is another story. The researchers themselves point this out. The drop in wanting therapy could mean people are feeling stronger and do not need as much help. Or it could mean people got comfortable enough with the AI that they stopped looking for the deeper help they might still need (Shoshani et al., 2026).

I do not know which story is true. The study cannot tell us. Both are possible.

I made myself sit with this. As someone who builds writing tools, the easy story is the first one. People feel better, so they need less. Done. But the second story is also possible. And the second story, if it is true, is something I would need to take seriously.

The thing the AI could not do

The study measured three kinds of mental health symptoms. Anxiety. Sadness. And something called PTSD, which stands for post-traumatic stress disorder, the kind of symptoms that come from living through something very hard or scary.

The AI helped with anxiety. The AI helped with sadness. The AI did not help with PTSD (Shoshani et al., 2026).

The researchers were honest about why. PTSD usually needs a special kind of treatment. It is not the kind of thing a daily chat can fix. It needs a trained person doing specific work with you, often over a long time.

This matters. It means the AI is good for some things and not for others. It means using the AI for the wrong kind of suffering would be like using a band-aid for a broken bone. The band-aid is not bad. It is just wrong for the job.

The bias check

I want to mention one more thing the readers of the study should know.

The lead author of the study and one other author have business connections to the company that makes the AI used in the study (Shoshani et al., 2026). They have shares in the company. One of them works there.

But it does mean the people who designed and ran the study had a reason to want it to work. That does not mean the study is wrong. The methods were careful. The numbers are real. Other researchers can repeat the work.

It does mean we should treat this as one strong study, not as the final word. Real progress in research takes many studies, by many teams, in many places.

The same way I have to watch my own bias as someone who builds writing tools. The researchers had to know readers would watch theirs.

The next chapter is about why I think the change is coming anyway.

References

Shoshani, A., Gurfinkel, B., Kor, A., et al. (2026). Efficacy of a Conversational AI Agent for Psychiatric Symptoms and Digital Therapeutic Alliance: A Randomized Clinical Trial. JAMA Network Open, 9(4), e266713.