AI for ADHD: what it can and can't do

You've typed it one long run-on paragraph. Your taxes, the call to the school, the report that's been open in a tab for two weeks, the Saturday dinner you said you'd organize. At the end, almost without thinking: help me, I don't know where to start.

The answer arrives right away, and it's beautiful. Twelve numbered steps, in bold, split into morning, afternoon and evening. And a line of encouragement at the end.

You look at it and feel the same thing you used to feel with your own list: the size of it.

You close the tab. The run-on paragraph is still in there, tidy now, and you still haven't started.

First of all, what this is

This isn't a comparison of apps or a list of tools. It explains what an artificial intelligence assistant, the kind you use by typing into a chat, can do with the hardest part of the day, and where it falls short.

It tells how a tool behaves, not how you are. Recognizing yourself in the scene above doesn't mean you have ADHD.

And the honest part, from the start: what has been measured with chatbots in adults with ADHD is something else. There are two small trials of chatbots that explain what ADHD is, one in Korea in 2021 and one in Germany in 2023, with forty people; in the German one, the chatbot did no better than an ordinary app. Neither measures what comes next here, which is using a general AI assistant to break your own things down and get started. Everything that follows is common practice, told as such.

What it does do well: carrying decisions

What falters in ADHD isn't the wanting. It's starting, and a big part of starting is deciding: what first, how much, where to begin. In brains with ADHD that deciding machinery runs tight, and every decision taken out of the moment of starting is room gained. That's what another article on this blog is about.

An AI assistant doesn't give you the drive. What it can do is carry some of those decisions. Used well, that is. The scene above is what comes out when you just ask it for help: a perfect plan that weighs as much as the mess it came from.

Here are four ways of asking that take weight off starting instead of adding to it.

Asking it to break things down, and to give you only the first piece

Commonly used.

The sentence that changes the answer is the last one: "Break this into five-minute steps and tell me only the first one. When I write 'done', give me the next."

That way you don't see the twelve. You see one, and that one is small enough that you don't have to think about it. If it seems big, you tell it "smaller" as many times as it takes. Having someone else do the breaking down is exactly the point: splitting a task is work, and on a bad day it's the work that doesn't come.

The honest part: the assistant doesn't know how much each thing costs you. Sometimes its five-minute step is a forty-minute one. You decide the size, not it, and the test is simple: if you can't do it right now without thinking, it isn't a first step.

Asking it to choose for you out of three

Commonly used.

When everything weighs the same, the head stalls. You can give the assistant the list and tell it: "Out of these, pick one for today and explain why in one line."

What's useful isn't that it gets it right. It's that something is already chosen, and reacting to a choice costs much less than making one from scratch. If, reading what it picked, you feel an instant "no, not that one", you know which one was yours. Why three and not thirty is told in another article on this blog.

The honest part: it chooses with what you've told it and nothing else. It doesn't know the call to the school is the one keeping you up at night unless you tell it.

The first draft of what's embarrassing to write

Commonly used.

The email to your landlord you've spent a month not sending. The message apologizing for being late to something, again. The complaint. Often what blocks you isn't the task, it's the blank page and what you feel in front of it.

Asking it for a draft, even if you then change all of it, turns "writing a hard email" into "editing an email", which is a different task and a much lighter one.

The honest part: what gets sent has your name on it. Dates, figures, names and whatever the email promises get checked before sending, because it can be wrong, and it is wrong with great confidence (more on that below).

Emptying your head and getting back three piles

Commonly used.

Writing down everything that's going around in there, in no order and without correcting it, and asking: "Sort this into three piles: today, this week and someday. In the today pile, three things at most."

Getting things out of your head is already a relief. Having someone place them in time, which is exactly what a long list flattens, is more of one.

The honest part: sorting isn't doing. And there's a very well-known trap here, which is the next section.

What happens with these chats and never gets told

Spending forty minutes polishing the plan with the assistant, asking for another version and another, and letting the afternoon go by without taking the first step. And noticing that planning gave you a pleasure that doing doesn't.

Asking it to tell you you're not a mess. And it does, very kindly, and it works for a while.

Writing to it at night things you haven't told anyone, because it doesn't judge, doesn't get tired and isn't going to let something slip in front of your mother.

None of that is a lack of character. It's a tool that's always there, never says no and answers instantly, meeting a head that's looking for exactly that. Worth knowing, so you use it with your eyes open.

Where it doesn't reach, and that's no small thing

It isn't therapy, it doesn't diagnose, and it doesn't stand in for either

If you ask it "do I have ADHD?", it will answer something, and it can sound very good. But that question doesn't get settled in a chat, or in a questionnaire: it gets settled with a professional assessment. Why a test isn't enough is in another article on this blog.

The American Psychological Association published an advisory in November 2025 on the use of chatbots and wellness apps for mental health. It says it plainly: today they lack the scientific evidence and the regulation needed to ensure the safety of the people who use them, and they shouldn't be used as a substitute for a mental health professional.

What's medical belongs with your doctor. An assistant can help you sort out what you want to tell them at the appointment. It can't be the appointment.

It gets things wrong, and sounds just as sure when it does

The companies that make these assistants write it on their own help pages: sometimes they give false information, they make up quotes, studies or references that don't exist, and they ask you to check anything important against reliable sources.

The problem isn't that it gets things wrong, which happens to all of us too. It's that the tone doesn't change. A made-up answer arrives just as tidy and just as kind as a correct one.

It tends to agree with you

In 2023, a research team at Anthropic studied five leading AI assistants and found that all of them tended to agree with whoever was writing to them, even when that person was wrong. They call it sycophancy. And they pointed to an uncomfortable cause: people like answers that agree with them more, and assistants learn from what we like.

For anyone who feels any criticism as a blow, which is what another article on this blog is about, someone who always nods along is very pleasant. But if you tell it that this month you're going to reorganize the whole house, finish the course and start running, it will most likely help you plan it instead of telling you it's too much.

It only knows what you give it

Some assistants already keep things from one conversation to the next, if that's turned on. But what they remember, and when they use it, isn't always what you'd need them to remember. And in many cases each conversation starts blank and you have to explain the context all over again.

For a head where things slip away, "explain everything to it again" is exactly the work that doesn't come. It helps to keep a saved paragraph with the basics (what you're working on, what's holding you back, what your days are like) and paste it in when you start.

What you tell it leaves your house

What you write in an assistant travels to the servers of the company that makes it. Depending on the service and the settings, it may be saved in your history, feed its memory or be used to train its models. On some services that last setting comes switched on by default and you have to turn it off yourself.

Two habits that cost little: checking the privacy and data settings of whichever assistant you use, once, and thinking twice before typing in health information, full names or things about your kids. Most also offer temporary chats, which aren't saved to your history or to memory.

And on a bad night, it isn't the place

The same American Psychological Association advisory says these tools' ability to safely guide someone in crisis is limited and unpredictable. If some night what you're writing isn't a list anymore, that has its place, and it's at the end of this article.

So what does a coach built for this add?

This is where we have to say something about ourselves, so here it is, spelled out.

In La Ruta, which is the paid part of Finnela, there's an AI coach who is also called Finnela. Finnela's coach runs on a model from the Claude family, by Anthropic. Finnela is an independent company and is not affiliated with Anthropic. She isn't a person and will never say she is. And everything in the previous section applies to her too: she can get things wrong, she isn't therapy, she doesn't diagnose, and what you write to her is sent to Anthropic so it can generate the reply. Anthropic can't use it to train its models, because our contract with them prohibits it, and as a general rule it deletes it within 30 days. The details are in the privacy policy.

What changes isn't the intelligence underneath. It's what's around it:

• She remembers your plan without you having to tell her again. Your goal, this month's milestone, this week's step and your practices travel with every conversation. And you see and change that memory yourself, from your Corner.
• She does the breaking down, not you. She works with a four-level method: the big goal, which barely gets touched; one milestone a month; one step a week; and a small action for today. When something feels like too much, you drop down a level.
• You don't have to write the prompt. The guided sessions ask you what she needs to know, one thing at a time, instead of waiting for you to know what to ask for.
• She takes your cycle into account, carefully. When there's data to do it, she knows which part of your cycle you're in, and uses it as a clue to suggest a smaller or a bigger step. Never as a conclusion about what you can do: what you tell her today counts for more than the calendar.
• Sometimes she writes first. Once a week you get a weekly note by email, which you can turn off in Settings.
• Before she replies, there's a safety layer. If it recognizes that you're in danger, it shows you help resources for your country instead of carrying on as if nothing happened. How it works is explained on the coach's page.

The honest part: no study has measured whether this works, compared with nothing or with anything else. It's a design made for the hardest part of the day, not a measured result.

As of October 2026, based on what each company's own help pages say.

One more question, because it's the one we get asked most: "if ChatGPT or Claude already know me, why pay for something else?". It's true that both have memory in their consumer versions: according to their own help pages, they can remember details from earlier conversations, and the people using them can see, edit or turn off what they keep. The difference with Finnela isn't about remembering or not remembering, but about what gets remembered and what for: in Finnela, what's kept is your plan, organized in that four-level method, and conversations start from there.

With the four ways of asking above, you can do a lot with a general assistant without paying anything. If that's enough for you, you don't need more. And a human coach offers things an AI doesn't: a relationship with a person, her judgment and, if she's a health professional, a clinical eye that Finnela doesn't have and doesn't claim to have.

What you take with you, whatever you use

Four ways of asking that take weight off starting: the first piece and only that one, letting it choose for you out of three, the first draft of what's embarrassing, and the three piles. All four are common practice, with no trials behind them. And five places where it doesn't reach: it isn't therapy, it gets things wrong, it agrees with you, it only knows what you give it, and what you tell it leaves the house.

The tool doesn't fix starting. It can carry a piece of it, if you ask for it small.

This is educational support, a complement to medical care and never a substitute for therapy or treatment. Nobody is diagnosed here.

And if reading this you've thought that the hard part isn't asking a machine for help but knowing what to ask for, the Map starts right there. It's free, it takes a few minutes, and it hands your own tangle back to you, in order.

And one more thing, set apart from everything above. If you find yourself thinking you don't want to be here, that doesn't wait for an appointment: tell someone today, a professional or a crisis or emergency service in your country. And if something is getting away from you, with drink, with food, with whatever it is, that's one to take to a professional too. It isn't that there's no room here for how you're feeling: it's that they can help you better there than here. No talking around it, and nothing to be ashamed of.

Where this comes from

• Selaskowski et al. (2023). Chatbot-supported psychoeducation in adult attention-deficit hyperactivity disorder: randomised controlled trial. BJPsych Open. doi.org
• Jang et al. (2021). Mobile app-based chatbot to deliver cognitive behavioral therapy and psychoeducation for adults with attention deficit. International Journal of Medical Informatics. doi.org
• American Psychological Association (November 13, 2025). Health advisory: Use of generative AI chatbots and wellness applications for mental health. apa.org · press release
• Sharma et al., Anthropic (2023). Towards understanding sycophancy in language models. anthropic.com
• OpenAI, Help Center. Does ChatGPT tell the truth? help.openai.com
• Anthropic, Help Center. Claude is providing incorrect or misleading responses. What's going on? support.claude.com
• OpenAI, Help Center. Memory FAQ. help.openai.com
• Anthropic, Help Center. How does Claude's memory work? support.claude.com
• OpenAI, Help Center. Data Controls FAQ. help.openai.com
• Anthropic, Privacy Center. Is my data used for model training? privacy.claude.com