AI-Deepfake Detection: Do You Trust ChatGPT More Than Peers?

AI-Deepfake Detection: Do You Trust ChatGPT More Than Peers?

Introduction: The surprising “trust gap” in spotting deepfakes

If you’ve ever wondered whether people treat ChatGPT like a shortcut brain for spotting fake news, new research from Fu & Hanaki (arXiv:2608.01540) digs into it—experimentally, not just with vibes.

The core question: Do people rely on ChatGPT more than their human peers to detect deepfake news? In a lab deepfake-detection task, participants had to estimate how much of an AI-generated news article was actually human-written. Then, they received advice from either ChatGPT (GPT-4) or human peers, and were asked to revise their answers.

The key headline result? Participants weighted ChatGPT’s advice much more heavily than their peers’ advice—at least in the main experimental waves. But the story doesn’t stop there: an added 2025 experiment brought in linguistic experts, and the relative reliance shifted again. That nuance matters if you’re thinking about real-world “AI detector” policies, classroom rules, or platform moderation.

Why This Matters: The “fight fire with fire” trade isn’t automatic

This research is significant right now because deepfake detection is getting operationalized faster than our understanding of human behavior catches up. Platforms and schools increasingly talk about detection tools—sometimes even embedding AI detectors or “AI writing detectors.” But here’s the tricky part: even if the detector is good, people might still use it poorly (or ignore it), depending on who they trust.

A concrete scenario you could apply today: imagine a news literacy workshop where participants try to judge authenticity. If you provide an “AI authenticity score” generated by GPT-4-style tools, this work suggests people may adjust their judgments more toward the AI tool than toward peer opinions—but only if the advice is actually high quality. In other words, the biggest lever may not be “who produced the advice,” but “how useful is the advice once people decide to listen.”

This study also builds on earlier AI-reliance research in a clean way. Past work often compared “algorithms vs humans” or “AI vs experts” in other tasks. Fu & Hanaki extend that by focusing on a case where AI is both the problem (deepfake generation) and the potential solution (AI-based detection advice). It’s the “dual role” problem: AI can mitigate AI risk—but only if humans don’t over-trust a weak tool or under-trust a strong one.

## What the researchers actually measured: reliance + performance in a judge–advisor task

To answer the trust question, the authors used a classic behavioral structure from economics and psychology called the judge–advisor paradigm. You can think of it like this: each participant is the “judge” who makes an initial estimate, then they receive “advice,” and finally they give an updated estimate.

The deepfake task: estimate the “human-written proportion”

Participants worked through 30 rounds. In each round, they read a synthetic Japanese “news article” that blended human and AI text. Their job was not simply “real vs fake.” Instead, they estimated a number:

• HMpro = the proportion of the article that is human-written
• from 0 to 100
  • 0 = totally fake (all AI)
  • 100 = totally real (all human)

This design is clever because it captures more uncertainty than binary labeling. It also mirrors real detector interfaces, which often output a score rather than a simple yes/no.

The advice treatments: GPT-4 vs peers (and later vs experts)

The main experiment used two advice sources:

Then, after seeing advice, participants submitted a second estimate.

How “reliance” was quantified: the Weight of Advice (WOA)

The authors measured reliance with WOA (“weight of advice”). Intuitively:
- if your final answer moves toward the advisor, WOA rises
- if you stick to your initial answer, WOA stays low

They also interpret WOA around 0.5 as a dividing line: above it, participants’ final responses are closer to the advice than to their own initial judgment.

How “detection quality” was evaluated: accuracy + advice quality proxy

Performance wasn’t only about final accuracy; they also measured improvement from initial to final.

They computed:
- initial accuracy
- final accuracy
- whether accuracy improved after advice

For advice quality, they used an “ex post proxy”:
- how accurate the advice itself turned out to be (even though participants didn’t observe that benchmark directly)

The big advantage: performance changes can be explained by both how much people rely and how good the advice is.

## Do people trust ChatGPT more than peers? Main experiment: yes, strongly

Now to the question you actually care about: Do people rely on ChatGPT more than peers to detect deepfake news?

The sample and setting

In the main experiment, the lab ran in 2023–2024, recruiting 87 participants (native Japanese speakers, all students at Osaka). They assigned:
- 42 to the Human treatment
- 45 to the AI treatment

Each person completed 30 rounds, meaning the final dataset effectively captured 2 identifications per round across the experiment (with some exclusions for WOA-definedness).

Reliance results: big WOA gap (AI > peers)

In the main experiment:
- Average WOA (AI treatment): 0.592
- Average WOA (Human treatment): 0.326

So participants didn’t just use AI advice “a little”—they weighted it much more. And importantly, the authors note that both means differ significantly from 0.5 (the benchmark point).

That directly supports the study’s first hypothesis: reliance is higher when advice comes from AI than from peers.

Performance results line up with advice quality

Participants’ final detection accuracy was higher under the AI treatment than under the Human treatment (statistically significant). But here’s the mechanism:

• The paper finds the overall performance differences are primarily driven by advice quality, not by the label “AI” itself.
• In plain English: GPT-4 advice was more accurate on average, and participants did better because they followed better guidance.

They also show the dependency clearly in their interaction analyses: AI advice helps most when advice quality is high. When advice quality is low, AI advice performs closer to (or not better than) peer advice.

What did not matter: how AI-heavy the article was

One might expect participants to rely more when the article is ambiguous—e.g., partially fake. The task included articles ranging from totally human to totally AI generated, plus mixed ones.

But the authors report that the human-written proportion (HMpro) did not systematically affect reliance or the decision to revise after advice. So the trust jump toward ChatGPT was not just “people are more confused, so they trust AI more.” Their reliance pattern was more stable.

## Why advice quality beats “AI-ness”: the interaction you shouldn’t ignore

This is the part that can save you from a bad policy mistake.

It’s tempting to conclude: “People trust ChatGPT more, so always use AI detectors.” But the study supports a more conditional claim:

Reliance determines whether advice can influence behavior; advice quality determines whether that influence improves performance.

A practical analogy: the “GPS depends on calibration”

Imagine you’re using a GPS app while driving. You might trust GPS more than a friend’s directions—that’s reliance. But if the GPS map data is wrong (low advice quality), your trust will lead you into trouble.

Fu & Hanaki’s results echo that “gated effect”:
- When participants heavily relied on advice, the effect of advice quality on improvement became much stronger.
- When participants didn’t rely, even good advice didn’t move the needle much.

The two-stage insight (activation vs integration)

As a robustness check (and to get deeper than WOA), the authors used an activation–integration model implemented via a Heckman selection correction. This separates:
1. Activation: whether participants choose to take the advice at all
2. Integration: how strongly they incorporate it after activation

Key finding:
- factors like advice source and beliefs influence both stages, but
- the advice–initial gap works in opposite directions:
- it encourages activation (you consider the advice more)
- but reduces integration (you hesitate to fully switch)

So advice doesn’t just “pull”; it interacts with how far it is from what you initially thought.

## When experts enter the room: reliance patterns can flip in 2025

The most interesting twist is that the AI>peers pattern from the main experiment wasn’t the whole story.

In 2025, the authors ran an additional study with the same overall procedure but:
- added linguistic experts
- shifted when the prior-belief survey was asked (to test timing effects)
- included a peer group drawn from earlier sessions (preHuman)

Advice sources in the additional experiment

The 2025 design used five conceptual advice pools across treatments:

They recruited 133 participants in 2025 and assigned:
- 44 to Expert
- 47 to preHuman
- 42 to AIadd
(plus separate cells for other replicated conditions)

Reliance order in 2025: experts > ChatGPT > peers

The WOA ordering in the additional experiment was:

WOA_AI > WOA_Expert > WOA_AIadd > WOA_Human ≈ WOA_preHuman

And crucially, when focusing on the clean within-wave comparison (Expert vs AIadd, both run under the same 2025 procedural flow), participants showed higher reliance on experts than on ChatGPT.

So the answer to “Do people rely more on ChatGPT than their peers?” depends on what peers and experts look like in the experiment—and on the time trend in how people think about AI detection.

## What beliefs do to reliance: trust is personal, but not purely about “daily ChatGPT use”

Beyond who gave the advice, the study also measured participants’ prior beliefs with survey questions about:
- whether they’d heard about ChatGPT
- how often they use it per week
- who (GAI vs humans) they think is more accurate at deepfake detection in that task

Then they built a variable capturing participants’ relative preference for the advice source: they were coded as preferring the offered advice if their belief matched it.

Two important patterns

1. People rely more on sources they prefer.
   Preference predicted reliance significantly. That supports the idea that reliance is “tied to trust,” not just to surface AI cues.

2. But daily usage of ChatGPT didn’t automatically translate into reliance.
   Frequency of ChatGPT use and belief in AI outperformance mattered differently in the way the authors analyzed reliance.

So even if you’re a heavy ChatGPT user, you might not behave like a “default AI believer” in a detection task.

## Key Takeaways

• In the main experiment (2023–2024), participants relied more on GPT-4 advice than on peer advice.
  Average WOA was 0.592 (AI) vs 0.326 (Human).

• Performance improved after advice in every group—but the advantage of AI advice came mainly from advice quality.
  In other words: better guidance led to better updates.

• Reliance matters because it gates the effect of advice quality.
  Good advice only helps when participants actually use it.

• The “AI > experts” story isn’t universal.
  In an additional 2025 experiment, participants relied more on linguistic experts than on ChatGPT when both were presented under a similar procedural flow.

• The amount of AI in the article (human-written proportion) didn’t meaningfully change reliance.
  People’s trust behavior didn’t scale straightforwardly with ambiguity in HMpro.

• Policy implication: pushing AI detectors isn’t enough. You need both:
  1) high objective detection quality, and
  2) public/user beliefs about which sources to trust.

If you want, I can also turn this into a “what this means for schools and platforms” practical checklist—focused on warning labels, detector integration, and how to avoid false-confidence effects from low-quality tools.