The AI Convergence Problem: Why Marketing Is Starting to Sound the Same

The AI Convergence Problem: Why Marketing Is Starting to Sound the Same

A Columbia and MIT study titled "The Basic Effect" found that LLM agents shift people's choices toward more popular options, reducing behavioral distinctiveness. A Science Advances study found that generative AI enhances individual creativity but reduces collective diversity of novel content across populations. Analysis of UK Parliament Hansard transcripts between 2007 and 2025 found AI-typical phrases spiked after ChatGPT's December 2022 release: "I rise to speak" reached a Z-score of 3.60 by 2025 against a 15-year stable baseline. And an Apple research paper titled "The Illusion of Thinking" found that frontier reasoning models experience complete accuracy collapse when puzzle complexity exceeds certain thresholds. The problem this creates for marketers: when every team uses the same AI tools on the same training data, outputs converge toward the same statistical center. Brand differentiation requires intentional divergence from that center.

Key takeaways

1. The Basic Effect (Columbia/MIT): LLM agents shift user choices toward more popular options, reducing behavioral distinctiveness. Marketing powered entirely by AI consensus will drift toward the industry average by construction.
2. Apple's "Illusion of Thinking" study: frontier reasoning models experience complete accuracy collapse when puzzle complexity exceeds certain thresholds, with performance drops of up to 65% on tasks where a non-answer-changing clause was added to math problems.
3. A Science Advances study found generative AI enhances individual creativity while reducing collective content diversity across populations: individual outputs get better, the population output converges.
4. The convergence formula is structural: shared data plus shared incentives plus fast iteration loops equals homogenized output, regardless of the team using the tool.
5. Three antidotes to convergence: proprietary asymmetric inputs (data or experiences competitors cannot replicate), visible human fingerprints (specific anecdotes, unconventional phrasing), and deliberately requesting opposite answers from AI to generate divergent options.

What the research shows

LLMs are, at their core, next-token predictors. Given a sequence, they predict the most probable next token. The statistical center of "most probable" is constructed from training data, which is drawn from the historical corpus of human text. That corpus is not neutral. It overrepresents popular choices, common frameworks, standard business vocabulary, and industry consensus.

When you ask an LLM to write a positioning statement, it predicts the most probable positioning statement given all the positioning statements it was trained on. When you ask it to write an email subject line, it predicts the most probable subject line for that context. The output is statistically average almost by design.

The ICP problem this creates for marketing teams using AI to generate brand content: competitive differentiation requires divergence from the statistical average. A positioning statement that is most probable given the training data is, by definition, the average of competitor positioning statements. Using it eliminates differentiation at the point of creation.

The named failure modes make this concrete: mode collapse (LLMs tend toward the same few completions despite multiple valid answers), the reversal curse (models trained on "A is B" struggle to apply "B is A"), and compositional collapse (performance degrades on novel combinations of concepts). These are not edge cases. They are structural tendencies that affect all current frontier models.

The car wash example from the research illustrates the consensus problem: when asked whether it was safe to walk to a car wash during operation, ChatGPT, Claude, and Grok all initially advised doing so (logically wrong consensus answer). Gemini and later Grok only arrived at the correct answer after it appeared in their training data. The models converged on a wrong answer because the wrong answer was statistically dominant.

Prooflytics builds marketing intelligence from tenant-specific accumulated knowledge: what your team tried, what worked, what did not, and what the market signal looks like for your specific accounts. This is a proprietary asymmetric input. The daily briefing is generated from signals that are not shared with any other tenant's model. The system is designed around the insight that generic AI outputs are cheap and converging, while account-specific intelligence compounds over time.

The LinkedIn MS Paint data point

A single LinkedIn post consisting of an MS Paint-style drawing generated 35,363 impressions, 448 reactions, 46 comments, and 24 reposts. This is not remarkable for its numbers. It is remarkable for what it demonstrates: visible human effort and obvious non-AI origin generated engagement rates that polished AI-generated content rarely matches.

The mechanism: AI-typical phrases (Navigating, Underscores, Streamline, Not just a X but a Y, Bustling, Is not merely) have become legibility signals for low-effort content. The UK Parliament data showing a Z-score of 3.60 for AI phrase frequency is not just a linguistic curiosity. It suggests that AI phrase patterns are now detectable to readers who have developed an implicit pattern-recognition for content origin.

For brand content: content that contains visible markers of human effort (specific personal anecdotes with named details, unconventional structural choices, concrete numerical claims from proprietary data) does not just avoid the convergence problem. It actively signals non-AI origin in ways that improve engagement.

Prooflytics

Make the call with the whole picture

Briefs are daily; the understanding compounds.

Start free trialSee pricing

14 days free · no credit card

What AI should and should not own in marketing

The research findings do not suggest eliminating AI from marketing. They suggest a division of ownership based on where convergence is a problem and where it is not.

Where convergence is acceptable (AI-suitable):

• Commodity execution: alt text, summaries, first-draft copy for standard formats
• Format compliance: adjusting copy to character limits, reformatting content for channels
• Research synthesis: aggregating information from multiple sources into a briefing
• Optimization testing: generating headline variants for A/B testing where statistical consensus is a starting point, not a final output

Where convergence is a strategic failure (human-required):

• Brand positioning statements and category definitions
• Campaign concepts and creative direction
• Competitive positioning claims
• Tone-of-voice decisions and vocabulary choices
• Audience framing and ICP definition

The article's formulation is direct: use AI for commodity tasks (alt text, summaries, drafting) and explicitly not for brand-defining decisions (positioning, headlines, concepts). The reasoning: brand-defining decisions are exactly where differentiation must occur, and AI consensus actively destroys differentiation.

01. Proprietary asymmetric inputs

The most durable antidote to AI convergence is feeding AI proprietary data that competitors cannot replicate. Categories of asymmetric input:

Customer language: verbatim language from customer interviews, support tickets, and review text. This is not available in AI training data in your specific form. Feeding exact customer quotes into a positioning prompt produces outputs that reflect actual customer vocabulary, not statistically average marketing vocabulary.

Internal performance data: your ROAS numbers, your conversion rates, your creative performance by format. This data is specific to your accounts. An AI briefing on "what worked in our last 30 days" is inherently differentiated from generic AI marketing advice.

Temporal specificity: recent proprietary market observations. What your sales team heard on calls last week. What a customer said was the tipping point for their decision. This is not in any training corpus.

Expert practitioner knowledge: specific professional experiences that are not documented publicly. A CMO who ran a specific campaign in a specific market with specific outcomes has information that is not available to a general-purpose LLM.

02. Visible human fingerprints

Four techniques that mark content as human-authored in ways readers respond to:

Named specificity: not "marketers often find that" but "Maria, head of growth at a B2B SaaS in Amsterdam, found that." Specificity is the primary marker that distinguishes observed experience from synthesized probability.

Unconventional structure: an essay that starts with the conclusion, a how-to that begins with failure mode before best practice, a comparison that starts with "both options are wrong." LLMs produce structurally conventional content because the training data is structurally conventional.

Proprietary numerical claims: specific data from internal analysis that is not reproducible from public sources. "Our analysis of 217 campaigns" is a stronger credibility signal than "research suggests."

Deliberate vocabulary defection: choosing words that are not in the AI phrase list. If the AI-typical phrase is "navigating the complex landscape of," the human alternative is a direct claim: "most teams mishandle this transition because they underestimate X."

03. Generating divergence from AI itself

A practical technique for using AI to escape its own convergence: prompt for the opposite, then work backward.

Step 1: ask the AI to generate the most generic, average version of the deliverable. This surfaces the statistical center.

Step 2: ask the AI to generate the most unconventional, contrarian version. This goes to the opposite statistical extreme.

Step 3: synthesize the two with proprietary asymmetric input to find a position that is differentiated from both the consensus and the deliberate contrarian.

This technique works because mode collapse produces the same output regardless of framing, but explicitly requesting the opposite forces the model to activate less-probable completions.

Bottom line

• The AI convergence problem is structural: shared training data plus shared tools plus optimization pressure toward consensus equals homogenized marketing output.
• The Basic Effect (Columbia/MIT) shows that AI-assisted choices drift toward the popular, reducing behavioral distinctiveness. For brand strategy, this is a differentiation risk that is invisible in standard performance metrics.
• The antidotes are also structural: proprietary asymmetric inputs (customer language, internal data, specific practitioner knowledge), visible human fingerprints (named specificity, unconventional structure, proprietary numbers), and deliberate divergence prompting.
• The division of labor: AI for commodity execution tasks, human judgment for brand-defining decisions.
• Prooflytics is built around proprietary asymmetric inputs: account-specific performance data, accumulated tenant-specific knowledge, and market signals that are not available to general-purpose AI. The intelligence compounds because it is specific, not because it is large.
• You can read independent reviews of Prooflytics on G2 and compare it to alternatives in the marketing analytics category.