AI note

The AI credibility crisis: what the evidence actually supports (internal synthesis)

Josh Shepherd7 min read2026-04-12

Executive summary

The underlying story Movemental wants to tell is broadly plausible: generative AI lowers the cost of fluent, polished, “credible-looking” text and imagery; platform incentives reward volume; and ordinary users are unsure how to authenticate what they see. But several specific numbers in circulation are misaligned with their likely sources or overstate what peer-reviewed work proves.

If Movemental wants a reputation for intellectual honesty (especially with research-minded movement leaders), the strongest public version of the argument will:

Anchor claims to named studies with scope statements (country, date, n, task type: discrimination vs. self-reported “fooled once” vs. workflow survey).
Separate “AI touched this document” from “AI wrote this document” from “AI replaced human judgment.”
Treat “credibility collapse” as a defined thesis, not an established academic term.
Pair alarm with counterevidence where it exists (e.g., top-ranking pages still disproportionately human-led in major SEO samples; expert annotators can detect many AI texts).

The “68%” statistic: likely a broken telephone

A careful pass for a statistic matching “68% of internet users struggle to distinguish human vs. AI content” did not turn up a primary source with that exact triad (68 / internet users / struggle).

What is well sourced:

Pew Research Center (U.S. adults, June 2025, n=5,023): 76% say it is extremely or very important to tell whether pictures, videos, and text are AI- or human-made; 53% are not too or not at all confident they can tell. That is a clean “importance vs. self-efficacy” gap, and it makes a credible argument without inventing a fraction.
All About Cookies (U.S. adults, February 2024, n=1,000, Prolific): 77% report they encountered something they believed was from a real person and later learned it was AI-generated. That supports pervasiveness + retrospective surprise, not a controlled measure of discrimination accuracy.

Hypothesis: “68%” may be a mangled recall of 76% (Pew importance), 77% (AAC fooled), or another adjacent figure from a secondary infographic.

Editorial recommendation: Delete “68%” until someone produces the original citation. Replace with Pew 53% / 76% (with link and date) or AAC 77% with explicit “U.S. self-report; not a lab discrimination task.”

“40–60% of blog posts are AI-generated or assisted”

This claim mixes three different empirical questions:

What fraction of published posts contain any model-generated wording?
What fraction are majority-AI with minimal human editing?
What fraction of professional workflows use AI for drafting or research?

The best large-scale prevalence proxy located for (1), with explicit limits, is Ahrefs’ April 2025 study of 900,000 newly detected English pages (one per domain) using an in-house detector: 74.2% contained some AI-flagged content, with only 2.5% labeled “pure AI” and 25.8% “pure human.” That is staggering, but it is not “blogs on major platforms” and it is detector-dependent.

Industry ranking studies (notably Semrush work summarized in Search Engine Land and Semrush’s blog) suggest pure AI pages underperform humans at position #1 in sampled SERPs. That complicates any straight-line story that “AI is already winning the credibility war everywhere.” A more accurate line: AI is flooding the long tail and middle of the web, while high-trust discovery slots remain contested, and may be more dependent on brand, citations, and human proof than before.

Recommendation: Retire “40–60%” unless tied to a specific report’s definitions. Prefer Ahrefs’ 74% contains some AI-detected text (new pages, English, Apr 2025) plus detector caveat, or cite workflow adoption separately.

Synthetic scale: real, but mind the superlative

NewsGuard and similar organizations have documented thousands of AI-dominated “news” domains and rapid growth. EU DisinfoLab traces coordinated inauthentic behavior at scale, including generative techniques in campaigns like Doppelganger.

Movemental should be cautious with “millions of synthetic accounts indistinguishable from humans” unless each clause is tied to a platform transparency report or a named measurement paper. The direction is right; the magnitude should not be freehand.

Can people detect AI? It depends which humans and which media

The cleanest “hard” result for multimodal synthetic content in challenging conditions is Cooke et al., “As Good As A Coin Toss” (CACM 2025; arXiv:2403.16760; DOI 10.1145/3729417): ~1,300 participants, realistic stimuli, mean accuracy near chance (~50%), with several factors that degrade performance further. The authors argue bluntly against relying on unaided human perception as a defense.

Text is not identical to video/audio. A Scientific Reports (2024) study on individual differences in text discrimination finds above-chance average performance but meaningful variance; fluid intelligence helps, and heavier social/smartphone use correlates with mislabeling AI as human (per reporting on DOI 10.1038/s41598-024-76218-y).

2025 ACL work shows frequent ChatGPT-using expert annotators can be very strong detectors under certain protocols. The nuance matters: expertise and task design matter.

Implication for Movemental: It is still fair to say most people, most of the time, should not bet their epistemology on vibe checks, especially as models improve. But it is unfair to imply nobody can detect AI or that all channels are already dominated undetectably.

Trajectories and “2–3 years”

Economic incentives (cheap generation, ad arbitrage, affiliate farms) plus tooling defaults (AI in docs, mail, social drafts) point to more AI-touch, not less. NewsGuard’s growth curves reinforce speed.

What we did not find: a consensus academic forecast that by 2028 “distinguishing real expertise from generated fluency will be nearly impossible without network verification.” That sentence is strategic projection, not a literature finding. It can still be used if labeled as forecast and paired with what verification must solve (adversarial adaptation, incentives to spoof graphs of trust, etc.).

Search, platforms, and expertise discoverability

Google’s public posture continues to evolve around spam, scaled low-value content, and quality, with third-party evidence that who ranks #1 in commercial queries still skews human-led in large samples. That does not negate AI’s impact; it sharpens it: the fight moves from “can you publish?” to “can you prove you were there, in relationship, in time?” That is actually friendly to Movemental’s deeper thesis about network-embodied credibility.

“Credibility collapse”

No standard paper titled credibility collapse was identified as a canonical term in this pass. Movemental can own the phrase if it defines:

Inputs collapsing: cheap signals (layout, tone, length, “researchiness”)
Verification lag: institutions still grading the old exam
Epistemic externalities: trust shifts to in-group attestations and dark patterns unless better infrastructure emerges

That is honest framing: coinage + definition, not faux-academic naming.

Religious / nonprofit / movement sector

Quantitative prevalence for “AI-written sermons” or “AI-drafted annual reports” was not established here. Qualitative evidence does show seminaries and theological educators wrestling with integrity, pedagogy, and policy (e.g., ATLA proceedings and theological librarianship literature in 2024). That is enough for a culture is catching up subplot, not for a fabricated sector percentage.

Strongest version of the argument the evidence currently supports

Self-efficacy crisis (U.S., high quality): Pew’s 53% not confident vs. 76% who say detection matters.
Material contamination of new web documents (industry measurement): Ahrefs 74.2% with some AI-detected content, with transparent detector limits.
Weaponized scale (investigative orgs): NewsGuard / EU DisinfoLab document industrial misuse, not anecdote.
Human limits under realism (peer review): Cooke et al. near-chance multimodal detection.
Non-collapse caveat (industry SEO): human-led pages still win many top slots, so Movemental’s offer is not “humans always lose” but rather: humans need better proof graphs to defend scarce attention.

Connection to Movemental’s constructive thesis

The research does not prove “scenius,” but it clears space for it:

If fluency is cheap, relationship-rich verification (who vouches, who co-labored, who saw the work happen) becomes more valuable, not less. This holds provided those graphs are harder to fake than text alone.
Movemental should be explicit that network verification is not a magic shield; it can be gamed by cliques and astroturf. The honest pitch is comparative advantage: harder to scale than paragraph farms; worth building anyway.

Counterarguments to keep in the margin

Detector-driven prevalence inflates “AI %" if detectors false-positive on certain human styles (non-native English, highly edited corporate tone).
Experts exist who detect well; some institutions will price discriminate toward them, creating inequality of discernment: a different social problem than universal fooled-ness.
Platforms may throttle the worst slop faster than pessimists expect, while still leaving mid-tier junk ubiquitous.

Institutional trust and “poll” context (Edelman, Pew)

Movemental sometimes needs a bridge from content authenticity to institutional trust. Edelman’s 2025 Trust Barometer materials (including AI-focused streams and flash polling) are useful for showing uneven comfort with corporate AI use and geographic divergence in AI trust, helpful nuance so Movemental does not imply a single global mood. Pew’s parallel finding that Americans are more concerned than excited about AI’s spread (50% more concerned than excited in the September 2025 reporting suite) complements the detection-confidence gap. Together, these sources support a societal unease story without requiring any single dramatic percentage about “fooled users.”

Verification systems and gaming (honest scope)

The prompt asks whether verification is being gamed at scale. The fair answer from open reporting is yes for some layers: SEO networks, affiliate templates, AI news farms, and coordinated operations adapt to rulesets quickly. Academic and industry literature also documents detector evasion and humanization tactics. None of that proves that every verification approach is hopeless; it does imply arms-race maintenance costs for any signal that stays purely syntactic (metadata badges, stylometry, cheap attestations). That is precisely why Movemental’s emphasis on relationship-backed verification needs to be articulated as economic and social hardness, not mystical certainty.

Closing note (tone)

This document is intentionally cool-headed. The marketing site can still be urgent, but urgency reads as maturity when the footnotes survive contact with Pew, arXiv, and the better industry studies. Movemental wins when it models the epistemic virtues it sells: scope discipline, sourced numbers, and a willingness to soften a claim rather than defend a rounded integer that slipped in from a press release chain.