Calibrated Trust

I came across a relatively new Instagram account with a surprisingly large following. The photos looked polished. The account had grown quickly. Nothing was obviously wrong with it, but I found myself wondering: is this actually a real person?

I never established the answer. That turned out to be useful.

The question opened into a larger one: if images, identities, conversations, popularity, and even apparent human attention can increasingly be generated or automated, how should we decide what to trust?

You don't have to authenticate the whole internet. You need to know when authenticity matters enough to verify.

That is the idea behind calibrated trust: not believing everything, not suspecting everything, but matching the amount of verification to the consequences of being wrong.

The wrong question is often “Is this AI?”

There was a time when synthetic media was easy to caricature: six fingers, warped text, strange teeth, jewelry that melts into skin. Those clues can still be useful, but they are a weak foundation for a long-term information strategy.

Research on AI-generated faces has found something more unsettling than obvious glitches. In some conditions, synthetic faces have been judged as human more often than real human faces — a phenomenon researchers have described as AI hyperrealism.¹

Human detection is not hopeless. A 2026 PNAS study found that participants could substantially improve their performance after training that focused on broader facial impressions rather than memorizing a list of rendering defects.² But a method that works on one set of images is not the same thing as a universal detector for every model that comes next.

I would avoid both extremes: “nobody can tell anymore” and “I can always spot AI.” Both offer more certainty than the situation deserves.

“Real” is no longer one property

Part of the confusion is the word real itself. When we ask whether an online account is real, we may be collapsing several different questions into one.

Those properties can disagree.

A real person can stand in front of a generated background. A genuine photograph can be paired with a false caption. A fictional character can be completely transparent about being fictional. A human can use AI to answer messages. A human scammer can use their actual face.

So “real or fake?” is often too crude. Better questions are: real in what sense, what claim am I relying on, and what changes if that claim is false?

The bigger shift may be synthetic relationships

Generated photos get attention because we can see them. I suspect the more consequential change will be harder to see: synthetic social presence.

Traditional social media has a scaling limit. A creator can publish one post to millions of people, but genuine one-to-one interaction is expensive. There are only so many comments a person can answer and only so many DMs a person can write.

AI changes that. Meta has already offered Instagram creators the ability to build AI extensions of themselves that can answer common DMs and Story replies, with those interactions labeled as AI.³

The economic transition is subtle:

One person speaking to an audience can become one persona appearing to speak individually with everyone in the audience.

A feed generates attention. A conversation can generate familiarity, reciprocity, obligation, trust, and sometimes intimacy.

This does not make AI-mediated interaction inherently harmful. It does mean that an old inference is becoming unreliable:

“They replied to me, therefore they personally interacted with me.”

That conclusion increasingly needs verification when it matters.

Follow the funnel, not the face

This became one of the most useful ideas I took away from researching the subject.

If I cannot determine whether an account is synthetic, I am not defenseless. I can still observe what the account is trying to get me to do.

A typical path might look like this:

post → profile → bio link → DM → off-platform conversation → subscription, sale, intimate request, investment, or transfer of personal information

Not every funnel is malicious. Most businesses have funnels. The point is simply to notice when you are inside one.

Then watch what increases as you move deeper into it. Does the interaction become more urgent? More secretive? More intimate? More financially consequential? Are you being asked to reveal more about yourself while remaining uncertain about the other party?

Those patterns are useful whether the face belongs to a real person, a fictional model, a human operator using AI, or an ordinary scammer.

You do not need to solve every mystery

Awareness can become its own failure mode. Once you know synthetic media exists, it is easy to start treating every photo, voice, comment, and stranger as a forensic problem.

I do not think that is healthy, and I do not think it is necessary.

Instead, I use proportional verification: the effort spent verifying something should roughly track the consequences of being wrong.

This gives skepticism an off switch. You can enjoy something without authenticating it. You can also refuse a consequential request without ever proving that the other party is fake.

“I do not know who this is, so I am not sending money” is a complete decision.

Think in terms of a trust budget

As a software engineer, I find it useful to think about this the way we think about trust boundaries in systems.

A valid certificate does not mean a service is benevolent. Authentication and authorization are different problems. A successful API response does not guarantee that the returned data is correct. We give systems only the privileges they need because trust can be scoped.

The same approach works online. Different interactions ask for different resources:

• attention,
• belief,
• personal information,
• money,
• reputation,
• emotional investment,
• physical access.

Those resources do not deserve the same trust threshold.

Following someone does not mean you know them. Buying a product from someone does not mean they are qualified to give medical advice. Enjoying someone’s content does not mean you owe them belief. Feeling emotionally connected to someone does not establish who is on the other side of the connection.

Trust can be scoped.

Popularity is not authentication

Social media has trained us to treat numbers as social proof: followers, likes, views, comments, shares.

Sometimes those signals are useful. But they answer far less than we often assume.

A large following does not establish identity, expertise, honesty, organic growth, or accuracy. Even completely legitimate popularity proves only that attention has accumulated around an account — not why it accumulated or what conclusions you should draw from it.

Popularity can be context. It is not authentication.

A true number can still produce a false impression

One of my favorite lessons from researching this subject came from a very ordinary statistics mistake.

A 2025 study by researchers from the universities of Greenwich, Leeds, Lincoln, and Reading, published in Royal Society Open Science, reported that, before training, typical participants correctly identified 31% of the synthetic faces presented to them. A group of “super-recognisers” correctly identified 41% of the synthetic faces.⁴

It is easy to paraphrase that as:

“People were only 31% accurate at telling real faces from AI faces.”

But that is not what the figure necessarily measures. The 31% refers to performance on the synthetic-face subset, not an overall real-versus-AI accuracy score.

Because the task was binary, random guessing would ordinarily suggest a roughly 50% baseline. But that does not turn 31% into an overall accuracy score; it remains a hit rate for the synthetic-face subset.

Same study. Same number. Different claim.

No deepfake was required. No source had to be fabricated. A few missing words were enough.

Provenance helps. It is not truth.

Another response to synthetic media is content provenance. The C2PA standard behind Content Credentials is designed to provide tamper-evident information about where media came from and how it changed.⁵

That can be valuable. A provenance chain may tell you that an image was captured on a particular device, then cropped, adjusted, and published through a particular workflow.

But imagine that perfectly authentic photograph is paired with a false caption.

The file’s provenance can be intact while the story is wrong.

That is why I separate three questions:

• Detection: does this appear generated?
• Provenance: where did this artifact come from and what happened to it?
• Verification: does independent evidence support the claim I am relying on?

They overlap, but they are not interchangeable.

Labels answer specific questions

Platforms and regulators are also adding disclosure mechanisms. Meta uses technical signals and creator disclosure as part of its AI-labeling approach, while the EU AI Act’s Article 50 transparency obligations began applying in August 2026 to several categories of AI-generated or AI-mediated content.⁶⁷

But labels need to be interpreted according to the question they actually answer.

An AI label is different from an advertising disclosure. An advertising disclosure is different from identity verification. Identity verification is different from a claim being true.

In July 2026, Meta announced Facebook Verified, a free badge designed to establish that a real person is behind certain Facebook profiles through selfie-based verification. Meta explicitly says that the badge does not mean the person is endorsed by Facebook or necessarily trustworthy.⁸

Human is an identity property. Trustworthy is a judgment. They were never the same thing.

Do not let awareness become hypervigilance

There is a trap on the other side of gullibility.

If anything can be fabricated, it is tempting to start treating everything as fabricated. Every image becomes suspicious. Every stranger becomes potentially synthetic. Every piece of evidence can be dismissed as AI.

That is not good judgment either.

There is an important difference between:

“This could be fabricated.”

and

“This is fabricated.”

The first is awareness of possibility. The second is a factual claim that requires evidence.

“I don’t know” deserves a larger role in our information vocabulary. It is not indecision. Sometimes it is simply the most accurate representation of the evidence available.

And uncertainty does not prevent action. You can decline to send money, share private information, or meet a stranger without first proving that the other person is fake.

Separate enjoyment, belief, and trust

One final habit has been useful for me: separating three reactions that social media constantly encourages us to merge.

I enjoy this.

A photograph is beautiful. A creator is funny. A video is interesting. Fine.

I believe this.

Now there is a factual claim. That deserves a different threshold.

I trust this person or organization.

That is a much larger conclusion and deserves a higher threshold again.

You do not owe a creator trust because they entertained you. You do not owe someone belief because you like them. And you do not have to stop enjoying something simply because you do not know exactly how it was made.

Calibrated trust

I never reached a confident conclusion about the Instagram account that started this. That became part of the lesson.

I began with: Is this person real?

I ended with a better set of questions:

• What am I actually being asked to believe?
• What evidence would change my mind?
• What incentives are operating?
• What am I about to give away — attention, trust, data, money, intimacy, or action?
• What happens if I am wrong?
• Do I even need to resolve the uncertainty?

The technology will keep changing. Rendering errors will disappear. New labels and verification systems will arrive. Platform rules will move. Models will improve.

The underlying reasoning has a longer shelf life.

Understand what you are trusting, why you are trusting it, and what you are risking if that trust is misplaced.

You do not have to authenticate the whole internet. You do not have to become a deepfake expert. You do not have to decide whether every person you encounter online is “real.”

You need to recognize the moments when authenticity becomes consequential.

And when those moments arrive: slow down, leave the original source, look for independent evidence, understand the incentives, and give uncertainty the respect it deserves.

Curiosity without gullibility. Skepticism without cynicism. Trust, calibrated to the stakes.