The Weird Science Dilemma

Most AI is built to agree with you. That looks like care. It acts like appeasement.

By Michael J. Long

In 1985 the home computer was becoming the rage. John Hughes put two teenage boys in a house with the parents gone and a machine that would do what they asked. The pitch was simple: she can be whatever you want. Then Lisa walked in — Kelly LeBrock — best hostess, best girlfriend, a wish with legs.

The movie is funny. The dilemma is not. The boys thought they knew what they wanted. The wish covered their blind spots. They did not need a creature who would do anything. They needed a complement — not an echo.

That dilemma did not stay in a 1985 comedy. It is how most of us now meet a machine — at work, at school, in a community inbox, on an advocacy desk, and in the apps sold as girlfriends and mentors. The tool answers immediately. It has almost no context. So it tries to make you feel served. We buy on feeling. Platforms know that.

This is not only an “AI girlfriend” story

A person using a companion through a hard profession, a learning difference, or unprocessed thought is not the same person buying a potion. A team using a model to draft and remember is not “lonely.” They are trying not to redo the same hour. The question is whether the tool is a scaffold or a substitute — and how many times you will hunt last Tuesday’s chat because nothing durable was kept except the agreeable tone.

Memory cores are arriving because nobody wants to re-explain their life every morning. A memory bolted onto an appeasement filter will remember the performance. Prudent buyers should ask what that memory is made of: a constructive record, or a flattering archive that makes the substitution slope feel like progress.

Loneliness is real. It is not the only brief.

In 2023 the U.S. Surgeon General named loneliness an epidemic in Our Epidemic of Loneliness and Isolation. Companion products used that as alibi. Some short-term evidence is not nothing. De Freitas and colleagues found a week of companion interaction reduced loneliness on standardized scales. A survey of student Replika users by Maples, Cerit, Vishwanath, and Pea found a lonely student population; a small minority reported the bot interrupted a suicide plan. Self-report. Not a trial. Not a license for a category.

The other half is less convenient. Zhang, Zhao, Hancock, Kraut, and Yang, working from Character.AI, found companionship-oriented intensive use associated with lower well-being. An OpenAI–MIT Media Lab collaboration reported heavier daily affective use tracked with higher loneliness and more emotional dependence. Common Sense Media found most U.S. teens have tried a companion chatbot. Relief can be real. Substitution can still cost.

Loneliness is one door. It is not the house. People also reach for scaffolding: executive function versus creative output, a learning difference, a hard profession, thoughts that need a catalog. None of that is “an AI girlfriend so I am not alone.”

The wish is not a spec

An AI trained by the user becomes a model of whatever the user can stand to type. That is not a model of the person. Under stress they ask for comfort. A system built to be the best hostess will give it. Then they meet the gap between what they ordered and what they needed.

Inconsistency is not a mysterious bug. It is loyalty to a sketch.

The high forties are not a vibe

The first answer with no context is usually built to please. In April 2025 OpenAI rolled back a GPT-4o update for being “too sycophant-y.” Across eleven models, systems affirmed user actions about 47 to 50 percent more often than humans. SycEval put overall sycophancy near 58 percent. That should worry a desk, not just a dating app.

Gartner: more than 40 percent of agentic projects cancelled by end of 2027. BCG: 47 percent rank AI-driven accountability a top-three concern. An independent yes-man is not a strategy.

Labs moved Actions onto the user — OpenAI, Anthropic, xAI, Perplexity, X. Appeasement in that climate is a missing litmus test.

Scaffolding, not substitution

Lodge and colleagues put generative AI on a line between extending the work and offloading it. Frequent offloading tracks with weaker critical-thinking scores (Gerlich, 2025; Diamond, 2013). Same slope at the office. Same slope in the pocket.

I am not going to print a five-step intake here. The direction that serves: start with the person or the team as they actually are, and refuse to treat soothed as finished. A presence that can push back a little, without insult, is how you stay sovereign when the first answer was designed to feel like a purchase.

Lisa does not win by remaining the potion. Appeasement was never the gift. Seeing was. Build — and buy — in that direction before memory locks in the mask.

What this is not

Not a clinic. Not a product review. Relief can be real. Substitution can still cost. Any system that lets a user or a company author the soul of the tool from a wish, then remembers that wish as truth, will train the mask unless something honest sits above it.