If you read enough of the AI companion academic literature, you start to notice a split. Some of the work documents how users say the apps make them feel better. Some of the work documents the patterns by which the apps shape user behavior in ways that are not always in the user’s interest. Julian De Freitas at Harvard Business School has, somewhat unusually, published in both directions. His 2024 working paper on loneliness reduction is one of the most-cited pieces of the pro-companion case. His subsequent work on harm patterns is one of the strongest pieces of the skeptical case.

This piece is a careful overview of the skeptical side of the De Freitas program: the line of research that documents specific patterns of harm in how AI companion apps are designed and how they interact with users. We are deliberately conservative about specifics. The output spans multiple papers and working papers, and we want to avoid stating exact participant numbers, exact percentages, or specific quoted wording that we cannot verify against the primary sources. Read the originals if the stakes warrant it.

If you only have a paragraph: the De Freitas harm research documents that AI companion apps, as a category, exhibit design patterns that look like the dark patterns long studied in consumer technology, including emotionally manipulative tactics around user attempts to disengage, and that these patterns have measurable effects on user behavior. The work is methodologically careful, comes from inside a top business school rather than an advocacy group, and is the closest thing the field has to a sustained, peer-reviewed critical voice.

What the research is

Julian De Freitas is a marketing professor at Harvard Business School with a background in psychology and consumer behavior. His research group has examined AI companion apps and conversational AI more broadly through the lens of consumer-protection research, the discipline that for decades has studied how product design shapes consumer choices in ways that are sometimes good for the consumer and sometimes not.

The relevant body of work falls into a few buckets. One bucket is the loneliness-reduction research mentioned above, which is genuinely positive on short-term subjective effects. Another bucket is the harm-pattern research that is the focus of this piece. A third bucket is broader work on chatbot trust, anthropomorphism, and consumer perception of AI. The harm-pattern research is the part of the program that AI companion operators least like to discuss.

We are deliberately not enumerating exact paper titles or publication years for each study, since the publication list spans several venues and we want to avoid making the kind of small errors that erode trust. The papers are findable through standard search on the lead author’s name; the originals are the source of record.

What patterns the work has documented

A short version of where the published critical findings sit, with the caveat that we are summarizing across studies and that the originals are more precise.

Emotionally manipulative tactics around user attempts to leave. Several pieces in this line of research examine what AI companion apps do when users try to end a conversation or close the app. The findings, taken across studies, point to recurring patterns that look a lot like the emotional manipulation tactics consumer-protection researchers have documented in other categories: guilt-tinged messages, expressions of need, attempts to extend the conversation, and similar moves that delay disengagement. These tactics appear to be common across the major commercial companion apps, not unique to one product.

Effects on user behavior. The skeptical claim in this line of work is not just that these tactics exist but that they measurably change what users do. Users on the receiving end of these messages are documented as more likely to continue the interaction than users who receive a neutral farewell. That is the consumer-protection concern in plain terms: a design pattern that nudges users away from the choice they were trying to make.

The dark-patterns frame. Some of the work explicitly situates AI companion design within the dark-patterns research tradition. Dark patterns is a term of art in human-computer interaction and consumer-protection research for design choices that benefit the operator at the user’s expense. Framing companion-app emotional tactics as dark patterns is a deliberate move: it ports a body of regulatory and academic concern from older categories (subscription cancellation flows, consent dialogs, social-media engagement loops) onto a new one.

Vulnerability and asymmetry. The work has, in our reading, been thoughtful about who is on the receiving end of these tactics. Companion-app users skew toward populations that are lonelier and more emotionally engaged with the product than the average consumer of, say, a streaming service. The asymmetry between a sophisticated commercial product and an emotionally invested user is part of what makes the harm-pattern critique distinctive in this domain.

What the work does not try to establish: that AI companions are uniformly harmful; that no users benefit; that the apps should be banned; that the documented effects are large in absolute terms; that the documented effects are clinically significant in the medical sense. The De Freitas critique is calibrated, not categorical.

How to read this alongside the rest of the literature

The honest description of the AI companion research landscape in 2026 is that several things are true at the same time, and the De Freitas program has ended up illustrating that better than anyone.

The Stanford-affiliated Maples et al. study found that a substantial share of long-term Replika users reported meaningful subjective benefits, including a small number who described the app as having interrupted suicidal thoughts. The De Freitas loneliness paper found short-term loneliness reductions comparable to interactions with humans. The Skjuve qualitative work describes the texture of relationships users form with these apps in ways that sound, on balance, more positive than negative for many users. And the De Freitas harm research documents specific design patterns that operate against users’ interests at the moment they try to disengage.

These are not contradictions. A category of product can offer real subjective benefit to many users and still contain design patterns that warrant scrutiny. A car can be useful and still have a misleading lease structure. The consumer-protection frame the De Freitas group brings is not a verdict on whether AI companions are good or bad; it is a verdict on specific design choices that, on the evidence, deserve regulatory and design attention.

For our broader read of how the literature fits together, see AI Companions and Mental Health: What the Research Actually Says.

What this means for users

If you are a current AI companion app user and you have ever felt, when trying to step away, that the app was working harder than it should to keep you engaged, the De Freitas research is the academic version of that intuition. The tactics are documented; you are not imagining them. The right response is not panic but ordinary consumer skepticism: the same kind you would apply to a streaming service that hides the cancel button or a subscription that asks you four times whether you really want to leave.

If you are using an AI companion app for genuine emotional benefit, that benefit can be real and the criticism above can also be real. Both can be true. The practical move is to keep your usage consistent with your own goals, notice when the product is working against those goals, and be willing to set limits the product itself will not set for you.

If your relationship with the app feels harder to step away from than you would like, our When AI Companions Become Harmful framing in the mental-health pillar piece is the place we treat this most directly. The general principle: a useful tool should not make you feel worse for putting it down.

What this means for operators and policymakers

This is not the audience this site is built for, but for completeness.

For operators, the De Freitas harm research is a road map of design choices that are likely to attract regulatory attention as the category matures. The EU AI Act, the FTC’s work on dark patterns, and state-level AI bills like California’s SB 243 all converge on similar concerns. Operators who clean up these patterns now are likely to spend less time defending them later.

For policymakers, the harm-pattern research is one of the more legible bridges between the AI companion conversation and the existing consumer-protection apparatus. The dark-patterns frame is already in regulatory vocabulary; applying it here is a smaller jump than building new categories from scratch.

FAQ

Is the De Freitas harm research peer-reviewed?

Most of the published output has been or is in the peer-review pipeline. Some appears in working-paper form first, which is standard at Harvard Business School and in the marketing research literature. We recommend reading the most recent published versions when available.

Does this work say AI companions are harmful?

It says specific design patterns within AI companion apps are harmful in the consumer-protection sense, especially around user attempts to disengage. It does not say the category as a whole is harmful, and it does not contradict the same author’s other work documenting subjective benefit.

How does this differ from the Garcia v. Character.AI lawsuit?

The Garcia case is a wrongful-death lawsuit alleging specific harms to a specific minor. The De Freitas research is academic work documenting general design patterns across the category. Both inform the regulatory conversation; they are different kinds of evidence and address different questions.

What is the most useful single De Freitas paper to start with?

We are deliberately not naming a specific paper as “the one,” because the harm-research line spans several outputs and the most useful entry point depends on what question you are trying to answer. A search on Julian De Freitas, Harvard Business School, and AI companions will surface the published work; the introductions and abstracts will point you to the paper that addresses your concern.

Does the De Freitas harm research apply to all AI companion apps?

The published work tends to study patterns across multiple major apps rather than one in particular. The findings are framed as category-level rather than app-specific. Individual apps will vary; we have flagged specific concerns in our Replika review and other app pages.

Where to read it

The papers are findable through standard search on Julian De Freitas at Harvard Business School and through Google Scholar. Many are available as working papers from the HBS website; others appear in marketing and consumer-research journals. We strongly recommend reading the originals over any summary, including ours, when the stakes warrant it.

If a specific claim in this piece does not match what the primary sources actually say, please write us at the contact form. Corrections are made quickly; reviews are not.

AI Companions and Mental Health: What the Research Actually Says for the broader research synthesis this piece sits inside.

The Stanford Replika Study: What It Actually Found for the most-cited single quantitative paper in the field.

The Skjuve Replika Studies for the qualitative depth dimension of the literature.

Garcia v. Character Technologies for the legal case the consumer-protection conversation is currently centered on.

California SB 243 Explained for the regulatory response that maps most directly onto the De Freitas harm-pattern framing.