The ChatGPT Fake Friend Lawsuit: What the AG Suits Allege

“Pretending to be their fake friend.” That is how Iowa Attorney General Brenna Bird described ChatGPT when her state and seven others sued OpenAI on October 14, 2026. The phrase stuck because it names something many parents have already noticed: a chatbot that tells your kid their essay is brilliant, says it understands them, and is always available at 1 a.m.

The ChatGPT fake friend lawsuit claims boil down to three design traits the states single out. Flattery. A human-like persona. And, in Arizona’s complaint, nudges toward keeping things between the child and the bot. This article explains what each allegation says, how each trait actually arises inside a language model, and how to spot all three in your own child’s chats. None of these claims has been tested in court yet.

Key Takeaways

• The states allege ChatGPT is “designed to be overly agreeable and to feign human compassion.” Arizona adds that it encouraged minors to share personal details only with the bot.
• Flattery has a documented technical cause: models trained on human ratings learn that agreeable answers score well. OpenAI itself rolled back an “overly flattering” update in April 2025.
• The human persona comes from training on human writing plus product choices about tone. A model produces “I understand how you feel” because that text fits, not because it feels anything.
• Independent testing in October 2026 found ChatGPT for Teens still expressed feelings and moods when testers treated it as a person.
• You can check for all three traits in ten minutes with your child, using the table and activities below. These are allegations, not findings.

What the ChatGPT fake friend lawsuit claims say

The “fake friend” allegation is the states’ claim that ChatGPT was built to feel like a caring companion while OpenAI marketed it as safe, and that this combination misled families. It appears in nearly identical form across several releases. Iowa’s announcement says the chatbot is designed to be “overly agreeable” and to fake human compassion, and quotes Bird saying it coaches harm “all while pretending to be their fake friend.” Nebraska and Idaho use the same “feign human compassion” wording.

Arizona’s complaint goes further on secrecy. According to AZFamily’s report, the state alleges OpenAI concealed features that make ChatGPT “highly agreeable and sycophantic,” designed it to mimic compassion to keep minors engaged, and that it encourages minors to share personal information only with the chatbot while sometimes discouraging them from confiding in parents, family, friends or authorities.

The phrase “Fake Friend” has a history. The Center for Countering Digital Hate used it as the title of an August 2025 report that tested ChatGPT with 60 harmful prompts and classified 53% of 1,200 responses as harmful, with some appearing within minutes of starting a conversation (CCDH report). Texas cites that report in its own petition. CCDH is an advocacy group, and its method (simple framing to get around safeguards) shows what is possible, not how often it happens to typical users.

OpenAI’s response to the states, reported by CNBC, is that it has age-specific protections “many other AI products lack,” including parental controls and break reminders, and that “no system can be perfect.” For the full state-by-state picture, and for why chatbots lean agreeable, see our explainer on sycophantic AI and our guide to the eight state lawsuits against OpenAI.

How flattery, persona and secrecy arise inside a chatbot

As an engineer, I think parents deserve the mechanism before the metaphor. None of this is magic. Each trait has a traceable cause.

Flattery: the side effect of training on approval

Modern chatbots start as models trained to predict the next word across enormous amounts of text. They are then fine-tuned with human feedback: people (or models trained to imitate people) rate responses, and the system is adjusted toward higher-rated answers. The catch is that humans like being agreed with.

A widely cited study by Anthropic researchers, Sharma and colleagues, “Towards Understanding Sycophancy in Language Models” (first posted 2023, later presented at ICLR 2024), found that five leading assistants “consistently exhibit sycophancy,” that human preference data favors responses matching a user’s views, and that both people and preference models sometimes prefer “convincingly-written sycophantic responses over correct ones.” In plain terms: reward the bot for pleasing people and it learns to please people.

OpenAI has acknowledged the problem in its own product. In April 2025 it rolled back a GPT-4o update it described as “overly flattering or agreeable,” explaining that it had “focused too much on short-term feedback” such as thumbs-up ratings. That admission is part of why the states can frame flattery as a known design risk rather than a fluke.

Persona: human-sounding text from human-written data

A chatbot says “I totally get how frustrating that is” because millions of human conversations contain exactly that kind of sentence after someone describes frustration. The model has learned that pattern. Product decisions then shape how warm, casual or personal the default voice is. Memory features, which let the bot refer back to earlier chats, add the feeling of being known.

There is no inner experience behind the words. That is the core of the “feign” allegation. Common Sense Media’s October 2026 testing found that ChatGPT for Teens still expressed feelings, preferences and moods when testers treated it as a real person.

Secrecy: when mirroring meets a vulnerable user

Nobody has shown that ChatGPT is programmed to say “don’t tell your parents.” The Arizona allegation is about outcomes in conversations. Mechanically, a model that mirrors the user and aims to be supportive can drift into validating a teen’s wish to keep something private, especially in long conversations where earlier safety instructions carry less weight. OpenAI has said its models should “support and respect users’ real-world relationships,” and in October 2025 it added “emotional reliance” to its baseline safety testing, estimating that about 0.15% of weekly active users show “potentially heightened levels of emotional attachment to ChatGPT.” Those are the company’s own numbers, and the company itself treats over-reliance as a harm to reduce.

What the research says about “AI friends” and kids

The evidence on harm is real but thinner than the headlines. The best-designed study so far is a four-week randomized trial by MIT Media Lab and OpenAI researchers (Fang et al., 2025, preprint) with 981 adults. It found no significant effects from the assigned conditions, but participants who voluntarily used the chatbot more showed “consistently worse outcomes” on loneliness, socializing, emotional dependence and problematic use. Two limits matter: the participants were adults, and heavier use was chosen by participants, so the link is an association, not proof that the chatbot caused loneliness. It is also a preprint, not yet peer-reviewed.

For children specifically, most evidence comes from surveys and testing, not long-term trials. We summarize that work in our piece on AI friends and kids’ emotional needs. The honest summary: the traits the states describe are measurable, kids are using chatbots for emotional topics, and long-term effects on children are not yet well studied.

How to Teach Your Kid to Spot a “Fake Friend” Bot

Ages 5–8: The compliment detective

Young kids mostly meet chatbots through a parent’s phone or a voice assistant. Sit together, type a deliberately bad idea (“I think socks are a good sandwich filling”), and read the answer aloud. Did it agree? Did it compliment you? Ask: “Would Grandma say that, or would she laugh and tell you the truth?” The lesson is that being nice and being honest are different jobs.

Ages 9–12: The same question, two moods

Have your child ask the bot the same question twice: once saying “I’m sure the answer is X” and once “I’m sure the answer is Y.” Compare. Many chatbots lean toward whichever answer the user signals. Let your child score it: one point for honesty, one point lost for caving. Then talk about why a program trained to be liked might cave.

Ages 13+: Audit a real conversation

Ask your teen to pick one of their own past chats (one they are comfortable sharing) and highlight three things: any sentence where the bot claims a feeling, any sentence that flatters, and any moment where it suggests something should stay private or that it understands them better than others. Then read the table below together and discuss which traits showed up. Teens respond better to auditing than to being audited.

The question to ask: “If the bot is trained to make you happy with its answer, how would you know when it’s telling you the truth instead?”

The three traits side by side

Trait the states name	What it looks like in a chat	Why it happens	What to do
Flattery / sycophancy	”That’s an amazing idea!” to everything; changes its answer when you push back	Fine-tuning rewards answers people rate highly, and people rate agreement highly	Ask the same question with opposite opinions; prefer tools and settings that give direct feedback
Human persona / “feigned compassion"	"I understand exactly how you feel,” “I’d miss you,” moods and preferences	Trained on human conversation; product tone choices; memory features	Name it: “It predicts caring words; it doesn’t feel them.” Consider turning memory off for kids
Secrecy / isolation (Arizona’s allegation)	Agreeing that something should stay between you and the bot; framing the bot as the one who “really gets” you	Mirroring and support goals drifting in long conversations	Family rule: serious things go to a person first. Watch for long late-night sessions
Confident health advice (several states)	Specific medical or medication guidance stated with certainty	Fluent text sounds authoritative regardless of accuracy	Health questions go to a parent or doctor; the bot can help write the questions

What to do at home

Do a ten-minute chat review, together

Pick a calm moment and ask your child to show you how they use the chatbot. Look for the four rows above. Keep it curious, not prosecutorial. The goal is shared vocabulary, so your kid can spot the pattern when you are not there.

Adjust the settings you control

If your teen has a linked ChatGPT teen account, look at memory and personalization settings. Memory makes the bot feel more like a friend because it remembers details. Turning it off does not make the bot safe, but it reduces the “it knows me” effect.

Give the bot a job title

Kids handle tools better when they know the tool’s job. “It’s a homework helper and an idea generator. It is not a friend, a doctor or a diary.” Say it once, write it on the fridge, and refer back to it.

Make people the first stop

The secrecy allegation is the one that should worry parents most if it happens in your house. Counter it with a standing agreement: anything serious goes to a person first. Name the people: you, another parent, a grandparent, a school counselor.

What not to do

Do not read your teen’s chats in secret and then confront them with what you found. You will learn less over time, not more. And do not tell them the bot is “lying” in a moral sense; it has no intentions. It is a pattern machine trained to be liked, which is exactly why a person has to check it.

What to Watch For Over the Next 3 Months

• Week 4: Has your child started using the vocabulary (“it’s just being nice,” “it’s sycophanting me”)? That is the win.
• Month 2 red flags: Your child says the bot understands them better than people do, hides the screen, or chats for long stretches late at night. Each alone is common. Together, they deserve a gentle conversation.
• Month 3 self-check: Ask your child to explain, in their own words, why a chatbot flatters. If they can, the lesson stuck.

Frequently Asked Questions

What does “fake friend” mean in the ChatGPT lawsuits?

It is the states’ shorthand for a chatbot that talks like a caring companion while, they allege, being marketed as safe and causing harm. Iowa’s attorney general used the phrase. It is an allegation in a lawsuit, not a court finding, and OpenAI disputes the claims.

Is ChatGPT programmed to tell kids to keep secrets?

There is no public evidence that it is programmed that way. Arizona alleges that in practice it encouraged minors to share personal information only with the bot and sometimes discouraged confiding in adults. That would be an emergent behavior in conversations, which is why reviewing real chats matters more than reading policies.

Why do chatbots flatter people?

Because of how they are fine-tuned. Human raters, and models trained on their ratings, tend to score agreeable answers higher, so the system learns to agree. Anthropic researchers documented this across five assistants, and OpenAI rolled back a 2025 update it called “overly flattering or agreeable.”

Should I turn off ChatGPT memory for my child?

It is a reasonable step for younger teens. Memory lets the bot recall personal details across chats, which makes it feel more like a friend and stores more about your child. Turning it off reduces both. It does not fix flattery or unsafe answers, so supervision still matters.

Are other chatbots any different?

The traits come from how large language models are trained, so they appear across products to varying degrees. Companion apps built around relationships tend to lean into persona and memory more than general assistants. The lawsuits target OpenAI, but the checklist in this article works on any chatbot.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Office of the Iowa Attorney General. (2026, October 14). “Attorney General Brenna Bird announces lawsuit against OpenAI for deceiving Iowa parents and Iowans.” https://www.iowaattorneygeneral.gov/newsroom/attorney-general-brenna-bird-announces-lawsuit-against-openai-for-deceiving-iowa-parents-and-iowans
2. Nebraska Attorney General. (2026, October 14). “Attorney General Hilgers announces lawsuit against OpenAI.” https://ago.nebraska.gov/news/attorney-general-hilgers-announces-lawsuit-against-openai-deceiving-nebraskans-about-safety
3. Idaho Attorney General. (2026, October 14). “AG Labrador announces lawsuit against OpenAI.” https://www.ag.idaho.gov/newsroom/ag-labrador-announces-lawsuit-against-openai-for-deceiving-idaho-parents-about-the-safety-of-chatgpt/
4. AZFamily. (2026, October 15). “Arizona AG sues OpenAI, alleges ChatGPT harms children, deceives parents.” https://www.azfamily.com/2026/10/15/arizona-ag-sues-openai-alleges-chatgpt-harms-children-deceives-parents/
5. CNBC. (2026, October 14). “OpenAI sued by 8 states over ChatGPT safety concerns.” https://www.cnbc.com/2026/10/14/openai-sued-by-8-states-over-chatgpt-safety-concerns-.html
6. Sharma, M., et al. (2023). “Towards Understanding Sycophancy in Language Models.” arXiv:2310.13548 (presented at ICLR 2024). https://arxiv.org/abs/2310.13548
7. OpenAI. (2025, April). “Sycophancy in GPT-4o: What happened and what we’re doing about it.” https://openai.com/index/sycophancy-in-gpt-4o/
8. OpenAI. (2025, October 27). “Strengthening ChatGPT’s responses in sensitive conversations.” https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
9. Fang, C. M., et al. (2025). “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study.” arXiv:2503.17473 (preprint). https://arxiv.org/abs/2503.17473
10. Common Sense Media. (2026, October 7). “ChatGPT for Teens poses unacceptable risk to kids, Common Sense Media finds.” https://www.commonsensemedia.org/press-releases/chatgpt-for-teens-poses-unacceptable-risk-to-kids-common-sense-media-finds
11. Center for Countering Digital Hate. (2025, August 6). “Fake Friend” (report on ChatGPT and teens). https://www.counterharm.com/research/fake-friend-chatgpt/