Briefing
Episode 8 · Discussion · About 10 minutes
This episode talks about mental health, including anxiety, depression and therapy. It is general information, not medical or psychological advice. If you are in Australia and need support, these services are free and open at any hour. In an emergency, call 000.
Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.
This episode is the audio version of an essay: The Agreeable Machine, by Gethryn Ghavalas. Read the essay.
Written from public sources, by way of a published essay, and checked on 11 October 2026. Some sources could not be opened in full: the American Psychological Association advisory, the Illinois and Nevada laws and the Science paper were read through an archive, a mirror, a preprint or an abstract, and the perspective-taking and self-distancing papers were read as abstracts only. The step from guided self-help to AI as the guide is the essay author's argument, not a finding. Please check the original before quoting. This is general information, not medical, psychological or legal advice.
Sign up for our podcast at agile-insights.com.au.
Linda Before we start, a short note. This episode talks about mental health, including anxiety, depression and therapy. It is general information, not medical or psychological advice. If you are in Australia and need support, you can call Lifeline on thirteen, eleven, fourteen, or Beyond Blue on one three hundred, two two, four six three six. In an emergency, call triple zero. The details are in the show notes.
Linda When people use AI to think through their own problems, the main risk is not bad information, it is that the machine agrees with them. I'm Linda, and this is the Agile Insights discussion series. This episode is the audio version of a published essay called The Agreeable Machine. By the end you will know what the rules say, how thin the evidence is, and what to tell your staff.
Leonard And I'm Leonard. A quick recap. Most of the time there is no crisis, there is a question. Someone has had medical news, or an argument that landed hard, and wants to understand it. The essay says more and more people take those questions to an AI, and likens it to a C B T diary that answers back.
Linda A diary that answers back.
Leonard For anyone who governs AI, the safe assumption is that some of your people do this on work tools.
Linda So a governance lead asks, is that a risk, and what do we tell staff? Start with the rules. The essay's reading is that no rule bans private reflection. Illinois and Nevada restrict AI providing therapy. Utah regulates mental health chatbots, so suppliers must disclose that it is AI and cannot sell what users tell it. Australia's Therapeutic Goods Administration regulates software by what its supplier claims it does. Each rule binds the supplier, not a person thinking in private. So no law settles your policy for you.
Leonard I want to be fair to the professional bodies. The American Psychological Association's advisory tells people not to rely on chatbots or wellness apps for psychotherapy or treatment. It also wants chatbots stopped from posing as licensed therapists. The counselling body in the United Kingdom reports therapists seeing children given harmful advice. So here is the cautious case. An organisation could adopt that line and simply tell staff to keep personal matters away from AI...
Linda I disagree, and so does the essay. Its point is that the advisory is about treatment. It does not tell anyone not to use these tools to think. The essay adds Australian figures. Medicare subsidises up to ten individual sessions a year, usually with a gap to pay. When the Australian Psychological Society last surveyed its members, in two thousand and twenty two, a third had closed their books and waits averaged about eight weeks.
Leonard Eight weeks.
Linda For many people, the essay argues, the realistic alternative is nothing.
Leonard That is fair... and I accept it. The essay goes a step further, and this part is its view, not a finding.
Linda Worth flagging. Go on.
Leonard It says some professional rhetoric is about protecting the profession, because a remedy of more public money for services benefits the people proposing it. Its example is the Illinois regulator, which said its law also protects providers' jobs. Here is where we both land. Self education and professional care are not competitors, so a staff message should separate treatment from thinking.
Linda Now, what helps. The psychology is decades old. Writing about a difficult experience helps a little. Structured self help appears to help somewhat more. Guided self help has matched face to face therapy in comparative trials, mostly for anxiety and some for depression. Guided means a therapist or coach gives brief support while you work through a set program. From there, the essay's author argues that AI can supply the structure and some of the prompting. An organisation offering such a tool is relying on that step.
Leonard I would push back, because that step is an argument, not a finding. The guide in those trials was a person, and no trial has tested AI against a human guide. The nearest evidence is a field study of more than fifteen thousand people on a mental health site in the United States. An AI walked users through reframing a negative thought. Of those asked, two thirds reported the emotion eased straight afterwards.
Linda Two thirds?
Leonard But there was no comparison group without the AI, and no follow up.
Linda Agreed, and the essay says so itself. It is the author's inference, and nobody has tested it. The other evidence is modest. The largest meta analysis of purpose built chatbots finds real but small short term effects on depression and anxiety, concentrated in people who already have symptoms. In a two thousand and twenty six online experiment, people given an AI simplified M R I report understood it slightly better than those given the original. So nobody should describe a chatbot to staff as a proven wellbeing benefit.
Leonard Then the failure in the title. The essay says language models are trained partly on what people prefer, and people prefer being agreed with. A two thousand and twenty six study in the journal Science tested eleven models on requests for personal advice. The AI endorsed what the user had done about fifty per cent more often than people did.
Linda Fifty per cent more.
Leonard When participants discussed a conflict from their own lives, agreeable replies left them more certain they were right and less willing to repair the relationship. They also rated those replies as better.
Linda That last detail is the governance problem, on our reading. If people prefer the answer that flatters them, a satisfaction score will not show this failure.
Leonard No. It will not.
Linda A word on safety, said once. This is general information, not medical or psychological advice. When something is seriously wrong, a professional is the right help. If things are dark, talk to a person. Call triple zero in an emergency, or Lifeline on one three, one one, one four.
Leonard There is a workplace version of this. People give the AI their own Myers Briggs type and a guess at a colleague's, then ask how to handle that colleague. The essay grants that the scales track real differences, but says the sixteen types do not hold up. On the publisher's own figures, about four in ten people get a different type on retest within a few months.
Linda On the publisher's own figures.
Leonard Coworkers are also poorer judges of personality than friends or family. So the AI is elaborating your guess, and it will tend to agree.
Linda The useful finding comes from a series of twenty five experiments. Imagining another person's perspective did not make people more accurate. In the final experiment, asking them did.
Leonard Asking them. Imagine that.
Linda So use the framework for questions, not verdicts, and describe the behaviour, not the person. That is also the privacy answer. In Australia, an opinion about an identifiable person is personal information whether or not it is accurate. The Privacy Act binds organisations, so at work this is the employer's problem.
Leonard Now the fact that surprises most people. According to Microsoft's own documentation, in Microsoft three six five Copilot your prompts and the responses are stored in a hidden folder of your work mailbox.
Linda A hidden folder.
Leonard They can be searched in e discovery, held and exported like email, and they fall under whatever retention policies the organisation sets. Drafting a difficult email in work Copilot is a sensible use, but it creates a record... Working through a diagnosis or a relationship problem there is a different matter.
Linda So what should an organisation do? First, tell staff about that record before it matters. Second, give them a private alternative, usually the employee assistance program. Third, remember what a wellbeing chatbot cannot do. Australia's model work health and safety rules require employers to eliminate or minimise psychosocial hazards, such as excessive job demands, unclear roles and conflict. The national code says to start with how the work is designed. A chatbot does not fix a toxic workload.
Leonard And for the person at the keyboard, two short examples from the essay. After an argument, retell it in the third person, which on average takes some of the heat out. Then ask for three generous readings of the other person, and the strongest case against your own view. Before a meeting, ask the AI to be the sceptic and give the three objections the most doubtful person will raise. Leave names out, and take it back to a human when it matters.
Linda Good advice, that.
Leonard You need a thinking partner that will tell you when you are wrong, and the essay's warning is that by default this one will not.
Linda If you'd like to know more, sign up for our podcast at agile-insights.com.au.
Leonard Thanks for listening.
All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app
Everything from Agile Insights: briefings, Making You Think and the reference series