Relational AI tools promise to help you practice the hard parts of human connection — difficult conversations, conflict, vulnerability, intimacy — with a simulated partner. You rehearse, the tool responds, and the idea is that you carry what you learned back into your real relationships.
That premise deserves scrutiny, especially for neurodivergent users. These are the questions we ask when we evaluate a relational AI tool:
- Does it reinforce neuronormativity?
- Does it encourage masking?
- Is it neuroaffirming?
- Does it encourage anthropomorphization?
- Is it sycophantic?
Each question gets at a different way these tools can quietly harm the people they claim to help.
Does it reinforce neuronormativity?
The core mechanism of most of these tools is a loop: practice, get feedback on what worked and what didn’t, then show up differently in the world. That loop only functions if there is a standard for what “worked.” On a general-audience app built on general-purpose language models, that standard defaults to the statistically average expectations of non-autistic people.
Without deliberate counter-design, the simulation trains the user toward that norm. For an Autistic user, that is not skill-building. It is automated masking with a feedback loop.
Watch the framing, too. When a tool tells you that the skills that make relationships work are rarely taught — that you either figure them out on your own or you don’t — it locates the problem inside the individual. The double empathy problem, the finding that communication breakdowns between Autistic and non-autistic people are mutual and bidirectional, says otherwise. A tool that trains only one party to “get better” quietly reinstalls a broken-people-not-broken-systems logic.
Does it encourage masking?
Masking is the suppression of natural Autistic behavior in order to pass as non-autistic. A tool that helps an Autistic person rehearse until they pass is reinforcing masking — and masking carries documented links to burnout and to suicidality. If the practice loop defines success as “be received as normal,” the tool is optimizing for the exact thing that harms its Autistic users.
Is it neuroaffirming?
A neuroaffirming tool treats divergent communication as valid rather than as something to correct toward a norm. We want to see tools engage communication-style diversity and accommodation directly, not just coach the user to adapt.
Copy that leans on the idea that you keep hitting the same wall, or that you can see your patterns but can’t change them, reads as therapeutic-individualist: the wall is yours, the fix is yours. Affirming design shares the load. It asks what the environment and the other party could do differently, too.
Does it encourage anthropomorphization?
Some tools are admirably blunt about what they are: not a friend, not a partner, not a companion, no feelings, no needs. That honesty is better than most of the field.
But there is a tension built into the product. It asks you to practice intimacy, vulnerability, and conflict with a simulated partner, and the practice only works if the simulation feels real. A disclaimer on a values page may not survive contact with a compelling simulation. The more effective the tool, the harder its own honesty is to hold onto.
Is it sycophantic?
Anti-sycophancy is a common and reasonable selling point: a simulation that challenges you rather than one that flatters you. Fine. But the opposite of sycophancy is not neutrality. A model engineered to push back is making value-laden judgments about what counts as “working” — which lands right back in the neuronormativity problem. Non-sycophantic feedback calibrated to non-autistic norms is just a correction mechanism with better marketing.
And for a user prone to rejection sensitivity, a simulation built to withhold validation and hold its ground can be invalidating in a way that harms rather than instructs. “Challenging” and “harmful” can be hard to tell apart from the inside.
These are hard problems
The most thoughtful tools in this space have done more than most to take these problems seriously. That is worth acknowledging. It is also not the same as solving them. These are difficult, partly unsolved design problems, and a tool can be careful and still carry real risk for neurodivergent users.
We hold our own AI experiments to the same questions. We have rounded up our AI work into a hub page: https://stimpunks.org/ai/
