Can Humans Tell? A Dual-Axis Study of Human Perception of LLM-Generated News

summary

The gist

Can humans tell whether a news article was written by a person or a large language model (LLM)? This paper investigates this question using JudgeGPT…

In short

Researchers used JudgeGPT to test if humans can reliably tell when a news article is written by a person or an AI model. Findings show humans cannot distinguish machine-generated text from human writing, regardless of the model size. This suggests user detection is ineffective, pointing toward system-level solutions like cryptographic content provenance.

Key concepts

JudgeGPT
A web platform used to collect human perception data on AI-generated text. It asks participants to rate fragments on source attribution (human vs. machine), authenticity (real vs. fake), and topic familiarity using continuous 0–100 sliders.
Dual-Axis Assessment
The method where participants rate news fragments across three independent continuous scales: source judgment, authenticity judgment, and topic familiarity. This approach allows for detailed parametric analysis of human perception.
Source Attribution
The judgment made by a participant regarding whether a piece of text originated from a human author or an artificial intelligence model. The study found participants could not reliably distinguish between these two sources.
Cryptographic Content Provenance
A system-level countermeasure suggested by the study. This involves using digital methods to track and verify the origin and history of digital content, ensuring users can trust where information comes from without relying on human detection.

Terminology used across episodes

This episode discusses

The paper

Can Humans Tell? A Dual-Axis Study of Human Perception of LLM-Generated News · Read on arXiv

Alexander Loth, Martin Kappes, Marc-Oliver Pahl

Frankfurt University of Applied Sciences · IMT Atlantique

Can humans tell whether a news article was written by a person or a large language model (LLM)? We investigate this question using JudgeGPT, a study platform that independently measures source attribution (human vs. machine) and authenticity judgment (legitimate vs. fake) on continuous scales. From 2,318 judgments collected from 1,054 participants across content generated by six LLMs, we report five findings: (1) participants cannot reliably distinguish machine-generated from human-written text (p >.05, Welch's t-test); (2) this inability holds across all tested models, including open-weight models with as few as 7B parameters; (3) self-reported domain expertise predicts judgment accuracy (r =.35, p <.001) whereas political orientation does not (r = -.10, n.s.); (4) clustering reveals distinct response strategies ("Skeptics" vs. "Believers"); and (5) accuracy degrades after approximately 30 sequential evaluations due to cognitive fatigue. The answer, in short, is no: humans cannot reliably tell. These results indicate that user-side detection is not a viable defense and motivate system-level countermeasures such as cryptographic content provenance.

DOI: 10.1145/3795513.3807431

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "Can Humans Tell? A Dual-Axis Study of Human Perception of LLM-Generated News".

Jane: Here's the gist of the paper. Can humans tell whether a news article was written by a person or a large language model (LLM)?

Tom: This paper investigates this question using JudgeGPT, which independently measures source attribution and authenticity judgment on continuous scales These results indicate that user-side detection is not a viable defense and motivate system-level countermeasures such as cryptographic content provenance.

Jane: Study Design and Methodology JudgeGPT is a web-based study platform designed to collect human perception data on AI-generated text The study employs a dual-axis assessment where participants provide three independent ratings on continuous 0–100 sliders: source judgment, authenticity judgment, and topic familiarity.

Tom: Stimuli are generated by RogueGPT, which orchestrates six LLMs with persona-based prompting to create machine-generated fragments.

Paper summary: Tom: So, to recap, the paper "Can Humans Tell? A Dual-Axis Study of Human Perception of LLM-Generated News" uses JudgeGPT to measure source attribution and authenticity judgment on continuous scales. The main thesis is that participants cannot reliably distinguish machine-generated text from human writing.

Jane: And they did this by collecting two thousand three hundred eighteen judgments from one thousand fifty-four people who rated fragments generated by six different large language models <ref:2604.03755#pg1>. They showed that the inability to tell holds across all tested models, even for open-weight models with just seven billion parameters <ref:2604.03755#pg1>.

Lu: The methodology involved a dual-axis assessment where participants rated each news fragment on three different continuous sliders: source judgment, authenticity judgment, and topic familiarity <ref:2604.03755#pg2>. They used these continuous scales instead of Likert items to really capture the degree of certainty people had.

Meng: And they made sure the stimuli were skewed toward machine-origin fragments—about ninety eight percent—to focus on how different AI models compare to each other rather than focusing just on human versus AI comparison <ref:2604.03755#pg2>.

Lalam: The results show that participants cannot reliably distinguish the two conditions, as seen in Figure two where the distributions for machine- versus human-written fragments overlap, with a non-significant t-test <ref:2604.03755#pg3>.

Tom: That's pretty clean data showing that no model was reliably identified as machine-generated across the board <ref:2604.03755#pg3>. It’s a big finding for anyone thinking about how we build detection features into applications.

Jane: And they didn't just stop there, though they looked at what makes people's judgments more accurate or less accurate, which is where the domain expertise part comes in.

Lu: They found that self-reported domain expertise predicted judgment accuracy with a correlation of point five three and a p value less than point zero zero one <ref:2604.03755#pg3>. That means if you know your stuff, you are more likely to judge the content correctly.

Meng: So, what about the other factor they looked at? They tested political orientation as well, and that didn't seem to matter for accuracy <ref:2604.03755#pg3>.

Jane: That’s right, political orientation showed no significant correlation with judgment accuracy; it was negative point one zero and non-significant <ref:2604.03755#pg3>. So, learned analytical skills matter more than ideological predisposition here.

Lalam: It also found that participants develop distinct response strategies, which they labeled "Skeptics" versus "Believers" <ref:2604.03755#pg6>. That’s interesting because it suggests how users might react to the content based on their existing disposition.

Tom: So to wrap up this segment, the paper establishes that human perception isn't reliable enough for user-side detection against these models, and it points toward domain expertise as a key predictor of accuracy. It sets us up nicely for thinking about what solutions look like beyond just asking users to spot the AI in front of them.

Jane: Right, so now we move from what they found to what it means for our industry and how we should be designing things moving forward.

Conclusion: Tom: We’re looking at the conclusion of this paper, "Can Humans Tell? A Dual-Axis Study of Human Perception of LLM-Generated News". The authors really drive home that the answer to whether humans can reliably tell is simply no.

Jane: They summarized their five findings in a way that confirms that user-side detection is not a viable defense against AI content generation <ref:2604.03755#pg1>.

Lu: The implication they are pushing for, which I think is the most important thing, is to move toward system-level countermeasures, like cryptographic content provenance <ref:2604.03755#pg1>.

Meng: From an engineering standpoint, that means we shouldn't rely on the user’s ability to spot AI; we need to build verifiable systems that prove the origin of a piece of content. That’s a different kind of challenge entirely.

Lalam: And they also mentioned that while the learning effect is there initially, accuracy degrades after about thirty sequential evaluations because of cognitive fatigue <ref:2604.03755#pg7>.

Tom: So we've talked about the lack of reliable human detection and the role of domain expertise as a predictor of judgment accuracy. The final message is that we need to focus our efforts on building these system-level countermeasures instead of trying to train people better at spotting AI content.

Jane: Exactly. So, the paper "Can Humans Tell? A Dual-Axis Study of Human Perception of LLM-Generated News" concludes that humans cannot reliably tell whether a news article was written by a person or an AI <ref:2604.03755#pg1>.

Lu: It’s a solid piece of research because it shows that the human perception side isn't the place to put our bets for defense against LLM-generated content <ref:2604.03755#pg1>.

Meng: We need to focus on building provenance tools, and Lalam, what do you think about that direction?

Lalam: I think cryptographic content provenance is the right direction because it’s a system solution rather than a user perception fix <ref:2604.03755#pg1>.

Tom: That sounds like a path forward for us in production, moving away from relying on users to be our first line of defense against AI content.

More episodes

← Home