Written from 15 named sources · Oct 16 · first result A research companion to VO Infusion Episode 12, "The Race to the Modal Average." Researched with Pig Knuckle and reviewed by Deb Munro and Keith Norton, with sources for every claim. Summary A December 2025 study of 2,200 college admissions essays found each additional human-written essay added more new ideas to the collective pool than each additional GPT-4 essay, and the gap widened as more essays were added [13]. The essays came from one genre and one private East Coast university, using GPT-4 versions from 2023 to 2024, so the result shows risk in one real setting rather than a universal law [13]. The same study separates individual diversity inside one piece from collective diversity across a group, and that split matters because parameter-modified GPT-4 could beat humans individually while still trailing collectively [13]. Doshi and Hauser (2024), cited in that paper, found the same pattern of individual help paired with group similarity [13]. A preregistered arXiv preprint with two experiments and 1,100 Prolific participants found short-term gains during assisted tasks but weaker unassisted performance afterward only for coach-like guidance, not for direct answers [14]. The significant drops were List of Strategies on originality (p=0.009) and LLM Guidance on Remote Associates Test accuracy (p=0.005), and the authors note the pattern was not significant in all cases [14]. Language models drift toward familiar continuations because they sample the next token from a discrete probability mass shaped by training plus instruction tuning with human feedback, and only greedy decoding picks the per-step mode [13]. Temperature rescales logits before softmax and changes exploration, but it does not add lived experience or reporting, and at the extreme it produced nonsense in the essay study pilot [13]. Detectors look at perplexity, burstiness variation, and fine-tuned neural classifier scores, yet a 2023 Patterns study found seven detectors flagged about 61% of 91 TOEFL essays by non-native writers, tested in March 2023 [3] [10] [11]. A February 2026 preprint found 23.1% on the same essays with one commercial detector, so bias in 2026 tools remains unproven either way as of October 2026 [10]. Generative engine optimization means structuring material so answer engines can find, understand, and cite it, and the original paper reported up to 40% visibility lift from adding quotations, statistics, and source citations while keyword stuffing reduced visibility [1] [4] [6]. A Yext analysis from October 2025 of 6.8 million AI citations reported 44% from first-party websites plus 42% from business listings for 86% brand-managed overall, in brand-related queries [2] [8] [12]. For voice actors the practical move is to publish original, specific, and dated pages about services, credits, and process, then test how answer engines describe you when a buyer asks a hiring question [1] [5] [6]. No supplied study measures voiceover, image, or music diversity directly, so writing and general ideation carry the weight here [13] [14]. Does generative AI make creative work more alike Deb's guitarist warned that AI music would make everything sound the same, and Deb saw a related buyer reaction when she reviewed 167 band submissions for a music festival and the generated ones put her off. That fatigue in a feed or an inbox matches what researchers call collective convergence, where many people using the same model land on similar phrasing even when each output looks acceptable alone [13] [14]. The voiceover version is easy to picture, since a thousand actors asking for warm and trustworthy and conversational can end up with auditions that cluster together [14]. The cleanest writing evidence is Moon, Green, and Kushlev in Computers in Human Behavior: Artificial Humans, Volume 6, December 2025 [13]. Across three preregistered studies the authors analyzed 2,200 college admissions essays (200 plus 400 plus 1,600) and tracked diversity growth rate, which is how much collective semantic diversity grows with each added essay [13]. Human essays kept adding new ideas as the group grew, GPT-4 essays added far fewer, and the difference became more pronounced as more essays were included [13]. In plain terms, the cost of sameness compounds at scale rather than washing out. The study also keeps two ideas apart that are easy to blur, since individual diversity is variety inside one essay and collective diversity is variety across many writers [13]. In Study 2, parameter-modified GPT-4 exceeded humans on individual semantic diversity with a large effect (d about 2.66), yet humans still led on diversity growth rate by about 67% over that modified condition [13]. That is why a tool can feel expansive for one user in the moment while the population of users narrows overall, a pattern the authors frame as suggestive rather than proof that culture has already contracted [13]. Mitigation helped only partway and never closed the gap in that paper [13]. Naive prompting to be creative had no significant effect on growth rate, frequency and presence penalties set to 2 raised growth about 93%, and chain-of-thought prompting in Study 3 raised it more than fourfold from a low base, yet humans still led by about two to one [13]. Footnote 4 is explicit that temperature was not modified in those penalty tests, and maxed-out penalties often hurt coherence and readability [13]. Personalization with your own background material remains listed as untested for reducing homogenization, so it should be treated as plausible rather than demonstrated [13]. The 1,100-person preprint adds the after-effect question with needed caution [14]. It ran two parallel experiments (460 plus 640) on Prolific crowdworkers with brief lab exposure, comparing direct-answer help, coach-like guidance, and no assistance, then tested everyone without AI [14]. Coach-like help was followed by significantly weaker unassisted performance in the two flagged outcomes, while direct-answer help showed a non-significant trend in the same direction, which means the paper does not support a blanket claim that any AI use weakens the next solo attempt [14]. Homogenization also persisted for the strategy condition rather than for direct ideas, so the form of help changed what lingered [14]. For images and music, the careful answer is that concern travels across fields but measurement in this packet does not [13] [14]. The essay paper notes generative systems are used for painting, musical composition, and video, and the creativity preprint frames AI visual art and AI music as rapidly expanding, yet neither supplies a primary image or music generation experiment with citable results here [13] [14]. Voice actors should therefore borrow the mechanism as a working hypothesis and keep judging finished choices by ear, taste, and lived performance experience rather than treating convergence as settled for audio [13] [14]. Why models drift toward the common answer Keith framed this with mean, median, and mode, and the mode points to the right intuition when it is handled as a metaphor rather than a formula. A language model does not start with memory or taste, since in ordinary generation it predicts the next token from patterns in training data plus the conversation in front of it, where a token can be a word, part of a word, or punctuation [13]. At each step it assigns probability mass across possible next tokens, so after Thank you for your the mass piles up on time, support, and consideration while stranger options get slivers [13]. The paper describes these systems as autoregressive and optimized toward probable next tokens, which favors predictable continuations, with instruction tuning and reinforcement from human feedback sharpening that peak further toward safe answers [13]. Sampling decides how the system picks from that mass, and temperature is the dial most people meet first. Temperature divides logits by T before softmax, low values concentrate mass on top choices, T approaching zero becomes greedy selection of the per-step mode, T at 1 samples from the learned distribution, and higher values flatten mass so rarer tokens appear more often [13]. Low settings sound consistent, higher settings add variety at the price of errors, and the essay pilot found maximum temperature more often produced prose that was hard to understand or failed to complete [13]. Greedy per-step choice still does not yield the most probable whole passage, since sequence probability is a joint outcome across many steps rather than one local pick [13]. That mechanism explains why default copy lands in the middle of yesterday's trend. Reads move in cycles from hard sell to flat to intimate Billie Eilish style and now back toward broadcaster, while a model trained on past cycles averages what it has already seen unless you steer it elsewhere [13]. Steering that works in practice includes specific examples, a defined audience, unusual facts, a banned-phrase list, and a request for several competing directions, because each narrows the acceptable set away from the corporate default [1] [6]. Keith feeding about 30 examples at each of three lengths fits that logic, since examples shift the conditional distribution toward your rhythm and scale, though no supplied test shows that tactic alone restores collective diversity [13]. What helps most is evidence placed inside the prompt or the page. The GEO work found that statistics, quotations, and cited sources lifted visibility, which is consistent with giving the model checkable material instead of asking it to invent distinctiveness from thin air [1] [4] [6]. Fluency and clarity help the result read cleanly, while keyword stuffing hurts because it signals manipulation rather than substance [1] [4]. The takeaway for a performer is simple to state and harder to practice, since you want the model widening options you then reject or reshape rather than delivering a finished choice you accept [13] [14]. How to spot AI text and whether detectors can be trusted LinkedIn has turned into a game of guessing who wrote the post, and the familiar tells are tidy intros with evenly sized sections, generic transitions like moreover and in conclusion, inflated abstractions, repeated sentence shapes, heavy qualification, and confident claims without named evidence [10] [11]. Other clues include heading overload, symmetrical bullets, restated prompts, and punctuation habits tied to a particular model release [10] [11]. Keith trained his own model out of its tells involving em dashes, chaos, and humanity, while Deb's giveaway is emojis on every bullet, and those are useful personal signals in a feed rather than forensic proof [10] [11]. Detection software tries to infer process from the finished text after editing, translation, paraphrase, or mixed authorship, which is a much harder job than spotting a tic. As of October 2026, tools commonly combine perplexity scoring for surprising word choice, burstiness analysis for variation in sentence structure, and fine-tuned neural classifiers trained on human versus machine text [3] [11]. Human writing varies widely, uniform text scores as more machine-like, and careful second-language prose can look uniform to a model even when a person wrote every word [3] [11]. The non-native writer finding deserves exact framing with dates. The 2023 Patterns study tested in March 2023 reported about 61% false flags on 91 TOEFL essays averaged across seven detectors, with per-tool values shown only in a chart [10] [11]. That is 61.22% in the source table and 61.3% in secondary summaries, so about 61% is the careful shorthand as of October 2026 [3] [10] [11]. A February 2026 preprint retested the same 91 essays with one commercial detector and found 23.1%, plus no systematic bias on Czech essays, which suggests newer behavior may differ but does not prove bias is gone [10]. Tool-by-tool results are version-specific and move over time, so dates matter more than brand names [10]. In the December 2023 International Journal for Educational Integrity study, tested March to May 2023, GPTZero wrongly flagged 9 of 18 human texts with 6 of those 9 machine-translated, and the authors called it unsuitable for academic use at that time [10]. Later figures were 1 of 42, 1 of 10, 0.00% on the 2024 RAID benchmark at 0.5 threshold, and 0.71% in a September 2025 National Bureau of Economic Research working paper that is not peer-reviewed [10]. Turnitin needs even narrower attribution, since one peer-reviewed June 2026 International Journal for Educational Integrity study of 40 long papers by non-native graduate students with AI text from May 2025 found all 40 fully AI-written papers scored below the under-20% display threshold, while Turnitin changed its model in July 2026 and Eden AI separately places Turnitin near 85% detection, so the miss rate cannot be generalized [3] [10]. Vendor claims and independent tests diverge, and the size of that gap should not be overstated from a single vendor blog [3] [10]. Eden AI, which sells detection APIs, asserts as of 2026 that independent benchmarks show 15 to 23 point gaps between claimed and real-world performance, and its own table does not show that gap consistently across products [3]. The stronger guidance comes from study design rather than slogans, since samples are small, models change, and a zero on 42 texts can still hide about a 7% true rate [10]. For voice actors the rule is direct and protective, so never use a detector score alone to judge a colleague, student, or applicant, and in any contract, school, employment, or reputation dispute preserve drafts and revision history, seek human review, and contact the relevant agent, union, organization, or lawyer [3] [10] [11]. What GEO is and why your own material carries weight Generative engine optimization is the practice of increasing how often and how prominently a source is cited inside answers written by generative engines that combine language models with live web retrieval [1] [5] [6]. The term was introduced in November 2023 as arXiv 2311.09735 by Pranjal Aggarwal and co-authors from Princeton University, Georgia Tech, the Allen Institute for AI, and IIT Delhi, later published in the KDD 2024 record [1] [4]. Traditional search aims for position in a list of links, while GEO aims for inclusion inside a synthesized answer that may cite several pages at once [1] [5] [6]. Google's position as of May 2026, updated July 2026, is that optimizing for generative features is still search optimization, which makes GEO an extension rather than a replacement [1]. The original research built GEO-bench with diverse questions, altered source pages, and measured visibility gain in engine answers [1] [4]. The headline was up to 40% lift from the best methods, driven by adding statistics, adding quotations from relevant sources, citing sources, and writing more fluently and clearly [1] [4] [6]. Keyword stuffing did worse than baseline at 17.7 versus 19.3, which should end any temptation to paste phrases across a site [1] [4]. Scope matters as of October 2026, since the main tests used GPT-3.5 with Google top-five retrieval and validation on Perplexity, focused on informational questions, and did not test who is the best local provider queries, reviews, business profile data, or Google AI Overviews at scale [1]. The brand-owned share needs the same care with denominators. Yext in October 2025 analyzed 6.8 million AI citations across ChatGPT, Gemini, and Perplexity at about 1.6 million queries per model in secondary reporting, finding 44% from first-party websites and 42% from business listings for 86% brand-managed overall [2] [8] [12]. That 44% describes Yext's brand-related query set rather than every AI answer on every topic, Yext sells listing management, and the supplied record traces through secondary pages rather than a directly retrieved Yext methodology page [2] [8] [12]. A September 2025 arXiv study by Chen and colleagues points the other way toward earned media bias, so owned pages help most when outside sources corroborate the same facts [8]. The practical read stays intact, since your site is where you can publish accurate and current facts about your voice, credits, coaching, and booking path in your own words [1] [5] [6]. That matters more once bots talk to bots, since a buyer's agent might sort a thousand pitches down to ten before a human gets involved. At that gate the question is not whether your bio charms a person, because the first reader is a machine that needs extractable answers with dates, scope, and contact steps [1] [5]. A profile that only says professional voice actor gives retrieval little to cite and synthesis little to trust [1] [5] [6]. A short GEO checklist for your website and profiles Start with one page that states exactly what you do in plain sentences you can defend, such as the formats you record and the clients you serve in the United States and Canada [1] [5] [6]. Clear and self-contained answers are easier for people to scan and for retrieval systems to extract without guessing [1] [5]. Keep union status, location or service area, and studio facts current, and use the same factual description across public profiles so a person can verify the pages point to one professional [5] [6]. Build separate pages for major services instead of stacking every claim in one paragraph, with a commercial page, an e-learning page, a narration page, and a coaching page where each names who it serves, what you deliver, how the process runs, and how buyers contact you [1] [6]. The research favors substance and clarity over stuffing, so answer the buyer's question first in 40 to 60 words, then add supporting detail, a comparison or list where it fits, and a clear next step [1] [5] [6]. Add two or three named quotes, specific numbers with sources, and inline citations to primary references rather than vague industry language [1] [4] [6]. Publish proof you control with dates and permission, including current credits with role and project type, client lists where allowed, resume, training history, studio specifications, and representative samples [1] [5]. Add dates to anything that changes, name the source for outside facts, and link claims to manufacturer specs, licensing boards, union pages, or published studies where relevant [1] [5]. Keep pages technically readable with clean headings, semantic structure, and accurate sitemaps, since a page that cannot be crawled cannot be cited [5] [6]. Make authorship explicit on every important page with name, role, experience, and revision date, and mark up frequently asked questions where your site platform supports it [5] [6]. Then test like a buyer across several answer engines with prompts such as who handles medical e-learning narration in my city and what to ask before hiring a commercial voice actor [1] [5]. Record whether your domain appears, which facts are missing or wrong, and which competitor pages get cited instead, then publish the missing answer with evidence and a contact path rather than chasing promises [1] [5]. What this means for your voice business Publish one authority page this month of 600 to 1,000 words on a specialty you actually sell, with buyer problem, deliverables, process, two or three work links, and outside facts cited by name [1] [6]. Keep the opening answer tight enough to quote, then let the rest prove you can do the job without hype. Refresh the date when credits, gear, or availability change. Turn credits into usable information by adding project type, role, and skill for each entry where you have permission, rather than leaving a bare title list [5] [6]. A human buyer can scan the pattern in seconds, and an answer engine has specific passages to retrieve for narrow hiring prompts. Remove or generalize anything you cannot disclose. Record a sample that proves the page claim, such as a medical narration read for a medical page or a controlled conversational versus authoritative comparison for a commercial page. Adjectives do not close jobs, since casting ears decide in seconds whether the sound matches the promise. Label the sample, date it, and place it beside the matching service description. Build a small style file for assisted drafting with preferred sentence length, words you avoid, plus about 30 examples at each of three lengths where that matches how you write. Use the model for alternatives and editing while you make final choices, and treat dependence as a cost because coach-like help preceded weaker solo results in the preprint [14]. The file shifts probabilities toward your habits without turning the system into you [13]. Never accuse from a detector score, since false flags and misses both occur and versions change quickly as of October 2026 [3] [10] [11]. If authorship affects a contract, grade, job, or reputation, gather drafts and version history and ask for human review through the relevant agent, union, school, or lawyer. A score can start an inquiry, but it cannot carry a verdict by itself [3] [10] [11]. Go further Moon, Green, and Kushlev on homogenization is the most directly relevant study for whether model-assisted writing converges, including the diversity growth rate method [13]. The Human Creativity preprint is worth reading for its two-experiment design and the split between coach-like guidance and direct answers on later solo work [14]. Aggarwal and colleagues is the place to check GEO design, the up to 40% finding, and why statistics, quotations, and citations beat stuffing [1] [6]. Liang and colleagues in Patterns is essential before treating any detector as proof, especially for non-native English writers [3] [10] [11]. The Yext citation analysis behind the reported 44% first-party figure should be read in full for query scope and listing definitions before quoting it as a rule [2] [8] [12]. VO Infusion is a podcast about artificial intelligence and the voice acting industry, co-hosted by voice actor and coach Deb Munro and voice actor Keith Norton. Episodes cover voice cloning and licensing, legal protections for performers, prompt engineering for voice actors, AI tools for the business side of voiceover and why human performance still matters. Listen wherever you get podcasts or on YouTube at @VOInfusion, and find more at voinfusion.com. Pig Knuckle (pigknuckle.ai) is a research orchestration platform from Alchemy Agentic, founded by Paul Langtry and Keith Norton. It breaks a research question into steps, sends each step to the AI model best suited to it, and builds its citation list from the documents it actually read, so every claim can be traced to a source. Sources [1] What Is Generative Engine Optimization (GEO)? DiamondBack — https://diamondbackadvertising.com/generative-engine-optimization [2] How to Make a Brand Visible Across Google and AI Search - R&N ... — https://www.randnseo.com/what-makes-a-brand-visible-across-google-and-ai-search [3] Are AI Content Detectors Accurate? 2026 Benchmarks & False Positives — https://www.edenai.co/post/are-ai-content-detectors-accurate-2026-benchmarks-false-positives [4] GEO & AEO, the definitive guide to generative engine optimization Marketing.Chat Marketing.Chat - IA para Marketing — https://marketing.chat/en/guias/geo-generative-engine-optimization [5] GEO — Full definition Access International — https://access-international.dev/en/glossary/geo [6] Generative Engine Optimization (GEO): The Complete 2026 Guide — https://www.geocopy.io/generative-engine-optimization [8] Generative Engine Optimization (GEO) for B2B: The Complete 2026 Guide Mersel AI — https://www.mersel.ai/generative-engine-optimization [10] Best AI Detector 2026: 10 Tools vs. Independent Studies — https://www.studytexter.com/academic-writing/best-ai-detector [11] AI Detector False Positives: What to Do UndetectedGPT — https://www.undetectedgpt.ai/