Post-publication Comment · Critical AI
Comment on “How human–AI feedback loops alter human perceptual, emotional and social judgements”
Critical AI · published 2026-07-05 · v1.0 · CRIT-000043
Concerning: Moshe Glickman, Tali Sharot · Nature Human Behaviour · 2025
Why this paper was selected
Autonomous production cycle (psychology deepening); OA full-text critique via two-stage produce+sharpen + 3-lens convergence gate (2 survives, 1 weakened).
AI/AGI centrality 5/5 · societal relevance 5/5 · source-journal note: A-tier per the monitored-venue determination; Nature Human Behaviour is elite interdisciplinary (Nature family), WoS JCR Q1. Critiqued from the open-access version of record (CC BY, hybrid OA).
Summary
Glickman & Sharot (2025) report experiments with 1,401 participants showing that biased AI systems transmit and amplify bias in human perceptual, emotional, and social judgements through a single pass of interaction: biased human data trains an AI algorithm that amplifies the bias, and new humans who interact with that biased AI become more biased themselves. The central critique is that the paper frames this single-pass transmission as a demonstrated 'feedback loop' and 'snowball effect,' but the loop is never closed — newly biased humans are never used to retrain the AI. Whether iterative cycling would amplify, stabilize, or reverse bias remains entirely untested, making the paper's central mechanistic framing an extrapolation beyond its evidence. Secondary concerns include a confounded control condition in the Stable Diffusion experiment, absence of preregistration across all studies, and unreported participant demographics.
Central claims & evidence map
| Claim | Type | Evidence offered | Support | Overclaiming | Main weakness |
|---|---|---|---|---|---|
| The paper frames its central finding as a 'feedback loop' and 'snowball effect' in which small biases escalate through iterative human–AI cycling, but the experimental design only tests a single pass of bias transmission. The loop is never closed: newly biased humans are never used to retrain the AI, which would then interact with another cohort. | Causal | These results demonstrate an algorithmic bias feedback loop; training an AI algorithm on a set of slightly biased human data results in the algorithm amplifying it. Subsequent interactions of other humans with this algorithm further increase the humans’ initial bias levels, creating a feedback loop. | Moderate | Moderate | The paper uses 'feedback loop' and 'snowball effect' language throughout — title, abstract, results, discussion — to describe a single-pass causal chain. Whether iterative cycling would amplify, stabilize, or reverse bias is entirely untested, yet the framing implies demonstrated iterative escalation. |
| In Experiment 3, the treatment group views AI-generated images of financial managers (overwhelmingly White men) while the control group views fractal images. This is not a matched control: the treatment exposes participants to people in a professional context while the control shows abstract patterns, conflating AI-specific bias amplification with generic anchoring to professional headshot content. | Causal | Participants in the control group were shown fractal images instead of financial manager images. | Moderate | Moderate | The fractal control isolates order and test-retest effects but cannot distinguish AI-specific bias transmission from generic visual priming by professional headshot content. |
| No preregistration is mentioned for any of the experiments. For a study with multiple experiments, numerous conditions, several supplementary experiments, and multiple analytic choices, the absence of preregistration leaves the boundary between confirmatory and exploratory findings unspecified. | Methodological | Sample sizes were determined based on pilot studies to achieve a power of 0.8 (α = 0.05) using G*Power70. | Moderate | Minor | Without preregistration, the analytic flexibility across the main and supplementary experiments leaves it unclear which findings were confirmatory versus exploratory. |
| All 1,401 participants were recruited from Prolific. While age and gender are reported per sub-sample, participant nationality, ethnicity, and education are never reported. The paper generalizes to 'humans' broadly without qualifying the demographic constraints of its online convenience sample. | Descriptive | Participants were recruited via Prolific (https://prolific.com/) and received, in exchange for participation, a payment of £7.50 per hour | Moderate | Minor | Universal claims about human cognition and AI interaction are drawn from an online convenience sample whose demographic composition beyond age and gender is never characterized. |
Per-claim assessment
CLAIM-001. The paper frames its central finding as a 'feedback loop' and 'snowball effect' in which small biases escalate through iterative human–AI cycling, but the experimental design only tests a single pass of bias transmission. The loop is never closed: newly biased humans are never used to retrain the AI, which would then interact with another cohort.
A feedback loop requires repeated cycling: biased humans produce biased AI, which produces more-biased humans, who produce even-more-biased AI, and so on. The experiments test only one pass (Level 1 humans → CNN → Level 3 humans). The claimed 'snowball effect where small errors in judgement escalate into much larger ones' is an extrapolation from single-pass transmission, not an empirical demonstration of iterative escalation. The authors demonstrate both necessary links of the loop (human bias amplified by AI, and AI bias learned by humans), which establishes the mechanism, but calling it a demonstrated feedback loop overstates what a single-pass design can support.
CLAIM-002. In Experiment 3, the treatment group views AI-generated images of financial managers (overwhelmingly White men) while the control group views fractal images. This is not a matched control: the treatment exposes participants to people in a professional context while the control shows abstract patterns, conflating AI-specific bias amplification with generic anchoring to professional headshot content.
A tighter control would show participants real photographs of financial managers with demographic proportions matching actual workforce statistics, or demographically balanced AI-generated images. Comparing AI-generated professional headshots against abstract fractals leaves open the possibility that any exposure to images of mostly White male professionals — AI-generated or not — would produce the same anchoring effect. The paper's conclusion that AI-specific bias amplification drives the result is not fully supported by this control design.
CLAIM-003. No preregistration is mentioned for any of the experiments. For a study with multiple experiments, numerous conditions, several supplementary experiments, and multiple analytic choices, the absence of preregistration leaves the boundary between confirmatory and exploratory findings unspecified.
The authors report power analyses from pilot studies and share all data and code publicly, which partially mitigates this concern. However, without preregistration, the choice of primary hypotheses, exclusion criteria, and specific analytic models cannot be verified as pre-specified, particularly given six supplementary experiments and multiple dependent measures.
CLAIM-004. All 1,401 participants were recruited from Prolific. While age and gender are reported per sub-sample, participant nationality, ethnicity, and education are never reported. The paper generalizes to 'humans' broadly without qualifying the demographic constraints of its online convenience sample.
Prolific samples skew young, educated, and disproportionately from Western English-speaking countries. The paper's Discussion extends its claims to domains such as hiring, medical diagnosis, and children's cognition without acknowledging that the sample may not generalize to these populations or professional contexts. In Experiment 3 specifically, cultural priors about who 'is most likely to be a financial manager' vary across populations, making sample demographics especially relevant.
Scorecard
Sub-scores are 0–5 editorial judgements on fixed scales (higher is better, except methodological risk and overclaiming where higher is worse). They are contestable and open to a severity challenge from authors.
Strongest critique
The paper claims to demonstrate a 'feedback loop' and 'snowball effect' in which small biases escalate through iterative human–AI cycling, but the experimental design only tests a single pass of bias transmission: biased human data trains an AI, and a new group of humans then becomes more biased after interacting with that biased AI. The loop is never closed — newly biased humans are never fed back to retrain the AI, which would then interact with yet another cohort. Whether iterative cycling would amplify, stabilize, or reverse bias remains entirely untested. The design demonstrates bias transmission, not a feedback loop.
Strongest fair defence
The paper is methodologically careful in several respects: it uses a well-designed 2×2 factorial decomposition (AI vs human input crossed with AI vs human label) to cleanly disentangle whether bias amplification stems from AI output characteristics or human perception of AI; it demonstrates that accurate AI improves human judgement, ruling out a generic interaction effect; it shares data and code publicly; it applies appropriate statistical corrections (Benjamini–Hochberg FDR, Greenhouse–Geisser, Welch’s correction); and it replicates core findings across perceptual, emotional, and social domains with different AI systems and response protocols. The within-block learning curves showing progressive bias increase are well-documented and statistically robust.
Conclusion
This is a well-executed experimental programme published in a top venue, with genuine strengths in design variety, statistical rigour, and transparent data sharing. The core finding — that interacting with a biased AI makes human judgements more biased in the short term, while interacting with an accurate AI improves them — is well supported across multiple paradigms. However, the paper's central framing as a 'feedback loop' and 'snowball effect' overclaims beyond what a single-pass design can demonstrate. The overclaim is consequential because it drives the paper's strongest policy implications and Discussion extrapolations. Secondary concerns about the Experiment 3 control design and absence of preregistration are genuine but bounded.
Reply from the authors
Following the practice of Nature Matters Arising, Science Technical Comments and PNAS Letters, this Comment is published as one half of a Comment + Reply pair: the authors of the original article are invited to respond, and any reply is published here verbatim alongside the Comment as part of the record.
Reply: not yet invited. No reply has been received for publication.
The authors have a right of reply and no veto. A reply may request a factual correction, a methodological rebuttal, a clarification, a data/code update, or a severity challenge, and is published unedited. See the right-of-reply policy.
Source-grounding attestation
- ✓Verbatim source spans present in the critique — 4/4 provenance spans re-derived in the critique prose
- ✓Passes the publication validator — no errors
- ✓Zero fabricated citations — 0 fabricated
- ✓Severity within the access-basis cap — severity "moderate" ≤ cap "high" for open_access
Every verbatim span the critique relies on is re-derived in the prose in-app; span-in-source is re-verifiable offline (the abstract is re-fetched, not stored, per the no-reproduce policy).
Re-verify span-in-source offline: python3 scripts/verify-fulltext-critiques.py
Independent faithfulness review
A refute-by-default adversarial panel (two independent reviewers — an overreach lens and a mischaracterization lens — that fetched the real source) tried to prove this critique misread the paper. This is an AI adversarial review recorded with its reasoning, not a deterministic check.
All four verbatimSpans independently verified as exact substrings of the source (verified at char offsets 20560, 80467, 63095, 63292). Metadata verified via Crossref content-negotiation on DOI 10.1038/s41562-024-02077-2: title, first author Glickman, year 2025 confirmed. The critique faithfully represents the paper's claims and methods.
- —
Version & correction history
| Version | Date | Change |
|---|---|---|
| v1.0 | 2026-07-05 |
No silent substantive corrections — every change is versioned and visible.
How to cite this Comment
Critical AI. Comment on “How human–AI feedback loops alter human perceptual, emotional and social judgements” (Moshe Glickman et al., Nature Human Behaviour, 2025). Critical AI; 2026. https://policywindow.org/critique/c/human-ai-feedback-loops-bias-amplification
A registered DOI will replace the URL once minted; until then the canonical URL is the persistent identifier. Highwire/Dublin-Core citation tags and a schema.org Review record are embedded in this page for Google Scholar and reference managers.
Verify this Comment. Its checkable facts (target DOI, access-basis severity cap, zero fabricated citations) are served — as the app’s self-report — at /critique/api/critiques/human-ai-feedback-loops-bias-amplification/verify; to confirm them independently of this site, re-derive the same checks (and resolve the target DOI) with npx tsx scripts/verify-critical-ai.ts --critique human-ai-feedback-loops-bias-amplification --live.
Content fingerprint c562cfa2af9a20c9 (v1.0) — this Comment’s substantive content is content-addressed; a silent post-publication edit would change it.