In February 2023, novelist Kazuo Ishiguro told The Guardian that he had experimented with GPT-4 to draft placeholder scenes during a period when he was blocked on a new project. He found the tool useful for generating structural scaffolding—rough scene beats, transitional paragraphs—but described the prose as "fluent without feeling anything." His observation crystallized a tension that writing researchers have since studied systematically: AI-generated literary fiction scores well on surface fluency metrics but poorly on what scholars call "affective specificity"—the pinpoint emotional detail that makes a moment feel witnessed rather than assembled.
The same year, a team at the University of Southern California's Information Sciences Institute ran a blind evaluation in which 240 readers rated short literary passages. Passages generated by GPT-4 with no human editing were rated significantly lower on "emotional authenticity" and "voice distinctiveness," even when rated comparably on grammar, coherence, and plot logic. The gap widened when passages depicted grief, shame, or ambivalence—interior states requiring what the researchers called "graduated uncertainty," a quality the model tended to resolve too neatly.
Literary fiction is distinguished from commercial genre fiction by its prioritization of interiority over event. The reader's primary experience is not what happens but how it feels to be inside a consciousness as things happen. This creates specific technical demands that stress-test AI writing systems in predictable ways.
Free indirect discourse is the technique by which a narrator temporarily inhabits a character's perspective without explicitly marking the shift—no "she thought," no quotation marks. Flaubert systematized it; Woolf and Joyce extended it into full-length streams of consciousness. AI models trained on large fiction corpora can produce surface-level approximations of free indirect discourse, but they tend to drift: the narrator's implied distance collapses or, conversely, becomes rigidly authorial when the scene calls for immersion.
Sentence rhythm as meaning is another pressure point. In literary prose, sentence length variations, mid-sentence interruptions, and the strategic use of fragments carry emotional information that supplements denotative meaning. A writer like Marilynne Robinson uses very long, syntactically complex sentences to enact a quality of sustained contemplation. AI prompts rarely specify rhythm constraints, and models default to a medium-length, declarative baseline that feels competent but not intentional.
AI models are trained on completed texts, not on the recursive drafting process. Literary prose is often shaped by what gets cut and revised across twenty drafts. The model's first output is structurally equivalent to a first draft by a writer who has read widely but has no personal stakes in this particular story.
Despite these limitations, professional fiction writers have documented real productivity gains from AI collaboration. The key is matching the tool to the task. AI performs reliably well on several fiction-adjacent functions:
Plot continuity checking. Submitting a draft chapter with a prompt such as "List every factual claim about character A's history made in this passage—flag anything that contradicts the following established facts" leverages the model's strong recall and pattern-matching against its own context window. Several authors in a 2024 survey by the Authors Guild reported using this method to catch timeline inconsistencies in novel manuscripts.
Generating rough dialogue variations. When a scene's dialogue feels flat, asking the AI to produce six alternative versions of the same exchange—each with a different subtext—gives the writer concrete material to react against. The writer rarely uses any version directly, but the variations clarify what tone and rhythm the scene actually needs.
Describing physical settings in multiple registers. Asking for the same location described once as a character in crisis would perceive it, once as a character in love, and once with the flat affect of clinical depression produces useful raw material. The writer then filters through their own voice and the specific character's history.
The single greatest risk in AI-assisted literary fiction is voice dilution—the gradual replacement of a writer's idiosyncratic style with the model's statistical median. This risk is not hypothetical. In 2023, a developmental editor at a major New York publisher described reading a manuscript that had clearly been AI-assisted because the prose "became grammatically cleaner but experientially flatter" midway through, then recovered its distinctiveness in the final third when the author apparently returned to unassisted writing.
Voice preservation requires explicit prompting strategy. Writers who report the best results use AI only after they have drafted a passage themselves, then ask the model to suggest edits "while preserving the following stylistic features"—listing specific traits: sentence length range, use of em-dashes, present-tense interjections, a particular vocabulary register. This "voice spec" approach forces the writer to articulate their style analytically, which has secondary value as a craft exercise independent of the AI output.
The most effective fiction writers using AI treat it as a reaction engine, not a drafting engine. They write something themselves, use AI to generate alternatives or variations, then react to the contrast. The writer's reaction—what they accept, reject, and modify—is where the literary voice lives.
In this lab you will explore how AI handles literary fiction constraints—free indirect discourse, affective specificity, and voice preservation. Describe a scene scenario or paste a short passage and ask the AI to generate variations. Experiment with giving it style specifications versus leaving them open. At least 3 exchanges to complete the lab.
In January 2023, CNET quietly published a series of personal finance explainer articles under the byline "CNET Money Staff" that had in fact been generated by an AI system. When technology journalist Maggie Harrison at Futurism investigated and published her findings in late January, CNET's editors acknowledged the program. A subsequent audit by CNET's editorial team found factual errors in at least 41 of the 77 AI-generated articles—including incorrect interest rate calculations, wrong loan repayment figures, and a misleading claim about FDIC insurance coverage. The episode became a widely cited case study in the journalism industry for the category of errors AI makes in factual financial and legal writing: errors that are structurally plausible but numerically or legally incorrect.
The CNET case revealed a specific failure mode: AI-generated factual prose sounds authoritative precisely because it mimics the style of authoritative sources. Readers—and editors—have to actively verify rather than rely on stylistic signals of reliability. Style and accuracy are decoupled in AI output in ways they typically are not in experienced human journalism.
Journalism operates under a set of accuracy obligations that differ from those of academic writing or creative nonfiction. In daily news reporting, the operative standard is verifiability: every factual claim must trace back to a named source, a documented record, or a directly observed event. Paraphrase and summary are permitted only when attributed to a specific human source or official document.
AI language models were not trained to satisfy this standard. They were trained to produce plausible continuations of text. When asked to summarize a policy, describe a historical event, or explain a technical topic, the model draws on statistical patterns in its training data—it does not query a database of verified sources. The result is prose that is often broadly accurate in its general claims but unreliable in specific details: dates, statistics, names, organizational affiliations, and legal specifics are the categories most prone to fabrication or distortion.
A 2023 analysis by the Reuters Institute for the Study of Journalism examined AI-generated summaries of 200 news articles and found that 38% contained at least one factual error when compared against the source articles, and 14% contained what the researchers classified as "significant" errors—those that materially changed the meaning of the reported information.
AI models can generate specific-sounding statistics, survey results, and institutional quotes with no underlying source. In journalism, a fabricated specific (a percentage, a dollar figure, a date) that sounds authoritative is more damaging than an acknowledged gap in knowledge.
Despite the accuracy risks, many working journalists and news organizations have identified specific workflow functions where AI assistance is both reliable and valuable:
Transcription and summarization of on-record material. When a journalist provides the AI with an actual transcript of a recorded interview or a public document, and asks it to identify key claims, inconsistencies, or thematic patterns, the model is working within its context window from provided text rather than from its training data. This reduces hallucination risk substantially. The Associated Press, which has used AI tools for automated financial and sports reporting since 2014, requires that the underlying structured data is provided to the model—not generated by it.
Translating technical language for general audiences. Regulatory filings, court documents, and scientific papers are written for specialist audiences. AI can produce multiple draft translations of technical passages at different reading levels, which journalists then verify against the source document and revise for their specific outlet's voice and context.
Identifying patterns across large document sets. Investigative journalists working with large FOIA document dumps, corporate filings, or legislative records have used AI to flag recurring terms, dates, or names for human follow-up. The AI acts as a first-pass pattern detector, not as a source of factual claims.
The fundamental incompatibility between AI output and journalistic sourcing standards is architectural: models cannot cite sources they cannot verify, and they have no mechanism for distinguishing between information they "know" with high confidence and information they have confabulated. The model's tone does not change when it produces false information—it produces false information in the same authoritative register as true information.
Working journalists who use AI productively have developed a discipline sometimes called the "source-first" rule: every factual claim in AI-generated or AI-assisted text must be traced back to a primary source before publication—not just checked against another secondary summary. This means treating AI output as a research starting point, not as citable content. The journalist must independently locate the original study, the official document, the direct quote.
In 2024, the Society of Professional Journalists updated its ethics code to include guidance on AI assistance, recommending that journalists disclose when AI tools have been used in reporting or writing, and affirming that the sourcing and verification obligations remain entirely with the human journalist regardless of what tools were used.
In journalism, AI is safest when working from provided text (a transcript you recorded, a document you obtained) rather than from its training data. Provide the source, ask the AI to work on that source, then verify its output against the source independently.
In this lab you'll practice identifying the verification requirements for AI-assisted journalism. You can provide a passage of text and ask the AI to identify which claims would require primary source verification, or ask it to translate technical language into accessible prose. Explore what kinds of tasks are safer to assign AI in a journalism context. At least 3 exchanges to complete the lab.
In 2022, Jasper AI (then Jarvis) reached a $1.5 billion valuation on the premise that AI could dramatically reduce the cost of marketing copy production. Within eighteen months, the company had undergone two rounds of layoffs and a significant strategic pivot. The story was not that AI copy tools failed—they succeeded at producing large volumes of serviceable text quickly. The story was that serviceable text at volume did not solve the marketing problem. Conversion rates on AI-generated copy without significant human revision were consistently lower than on strategically crafted human copy, particularly for high-stakes conversion points like landing pages, pricing pages, and email subject lines.
A 2023 analysis by Wynter, a B2B message testing platform, tested identical products marketed with AI-generated copy versus human-revised copy across 1,400 target-audience respondents. AI-generated copy scored 23% lower on "resonance"—the degree to which readers felt the copy understood their specific problem—and 31% lower on "distinctiveness"—the degree to which the copy felt different from competitors. The human-revised versions maintained these advantages even when the human revision involved only thirty minutes of editing on an AI draft.
Marketing copy is a genre defined by competitive differentiation. Its function is to make a specific product or service feel meaningfully distinct from alternatives in the mind of a specific reader. This requirement is structurally at odds with how language models generate text: by producing statistically likely continuations of the input. The most statistically likely continuation of marketing-register text is also the most common marketing-register text—which is to say, the clichés of the genre.
Professional copywriters have documented a predictable pattern in AI-generated marketing copy: benefit claims are generic ("streamline your workflow," "save time and money," "empower your team"); emotional appeals are safe and shallow ("imagine a world where…," "you deserve…"); and calls to action are formulaic ("get started today," "learn more," "try it free"). None of these phrases are technically wrong—they are used constantly in marketing because they're familiar—but they carry no differentiation signal. They are statistically correct and strategically neutral.
This is a genre-specific version of a broader AI writing pattern: the model optimizes for not being wrong rather than for being memorably right. Effective persuasive copy often requires being surprising, even jarring—violating reader expectations in a way that creates attention and recall. Models trained to avoid surprising outputs are structurally disadvantaged at this task.
The worst marketing clichés are also the most statistically prevalent marketing phrases in training data. The more a phrase has been used in marketing copy, the more likely AI is to generate it. This inverts the strategic logic of differentiation.
Despite differentiation failures, AI tools have proven genuinely valuable in specific marketing copy functions where volume, variation, and structural consistency matter more than creative distinctiveness:
Generating variation sets for A/B testing. Given a single human-written headline or subject line that performs well, AI can produce twenty structurally similar variations in seconds. Human writers then evaluate the variations against the brand voice and competitive context. This dramatically speeds up A/B test setup without requiring AI to generate the strategically strong original.
Adapting copy across channels and formats. A long-form product description can be compressed into a social media caption, an email preview line, a meta description, and a Google Ads headline, each with format-specific length and tone constraints. These are structural transformations that AI performs reliably well when given clear format specifications.
Localization and audience segmentation drafts. A single piece of copy can be adapted for different audience segments—different industries, different buyer personas, different geographic markets—producing raw drafts that human marketers then verify for cultural accuracy and brand consistency. HubSpot, Salesforce, and Mailchimp all reported in 2023 that their content teams used AI for this function extensively.
The most effective professional copywriters working with AI have developed what the industry now calls a voice brief: a structured document that specifies tone, vocabulary range, forbidden phrases and clichés, target reader's primary anxiety or aspiration, and two or three example sentences that exemplify the desired register. The voice brief is submitted with every AI prompt as system context.
In 2024, the Content Marketing Institute surveyed 312 marketing content professionals and found that teams using a documented voice brief with AI tools reported a 47% reduction in the number of revision rounds needed to bring AI-generated copy to publishable quality. Teams without a voice brief reported that AI copy required more total editing time than writing from scratch—because the editor had to first identify and remove the model's defaults before building toward something distinctive.
The voice brief approach also has strategic value independent of AI: it forces the marketing team to explicitly define what makes their brand voice different, which many teams have never done formally. In this sense, AI has created pressure to formalize brand language standards that should have existed but often did not.
Use AI to generate variations of human-authored originals, not to produce the originals themselves. The strategic insight—what makes this product genuinely different, what this specific reader most fears and wants—must come from human understanding of the market. AI's role is efficient variation and structural transformation, not strategic differentiation.
In this lab you will practice building a voice brief and using it to constrain AI marketing copy, then compare results with and without the brief. Explore differentiation, A/B variation generation, and channel adaptation. At least 3 exchanges to complete the lab.
In June 2023, a New York federal court sanctioned attorneys Steven Schwartz and Peter LoDuca of Levidow, Levidow & Oberman for submitting a brief that cited six non-existent court cases generated by ChatGPT. Schwartz had asked ChatGPT to find supporting cases for a personal injury lawsuit against Avianca Airlines; the model produced case names, docket numbers, and detailed judicial reasoning that sounded authoritative and professionally formatted. When opposing counsel was unable to locate the cases, Schwartz asked ChatGPT to confirm their existence—which it did—before the fabrication was uncovered. Judge P. Kevin Castel imposed sanctions and required completion of continuing legal education, noting that a cursory check of Westlaw or LexisNexis would have immediately revealed the citations as non-existent.
The case became an immediate landmark in professional training programs across law, medicine, and academic research. Its significance was not that an attorney misused AI carelessly—it was that the AI's fabricated citations were indistinguishable in style and format from real ones. The system produced false information using the precise typographic, citation, and jurisdictional conventions of real legal writing. This is the defining risk of AI in technical and academic genres: the genre's own conventions of precision and formality become camouflage for fabricated content.
Technical writing is a genre defined by two often-conflated qualities: technical register (the vocabulary, structure, and conventions of a specialized domain) and technical accuracy (the factual correctness of the claims made within that register). AI models are very good at producing technical register and unreliable at guaranteeing technical accuracy.
This creates a particular trap: technically formatted content reads as authoritative to non-specialists, and the conventions of technical writing—passive voice, hedged quantitative claims, systematic section structure, citation formatting—function as trust signals that readers use to assess reliability. When an AI produces technically formatted content that is factually incorrect, it borrows the trust signals of the genre to smuggle in misinformation.
A 2024 study published in Nature found that AI-generated scientific abstracts were rated as credible by non-specialist readers 69% of the time, compared to 68% for human-written abstracts—an essentially indistinguishable rate. However, the AI-generated abstracts contained methodological errors that specialist reviewers identified at a rate of 42%, compared to 9% for human-written abstracts. The genre's conventions successfully masked the errors from non-specialists.
In technical and academic writing, AI uses the genre's own precision conventions—citation formatting, quantitative hedging, systematic structure—as camouflage for fabricated or inaccurate content. The more formally correct the output looks, the more dangerous the unchecked error becomes.
Despite significant accuracy risks, technical writers and researchers have documented several workflow functions where AI is genuinely productive and where the risks are manageable:
Plain-language translation of verified technical content. When a subject-matter expert provides verified technical documentation and asks the AI to produce a version for a non-specialist audience—specifying reading level, vocabulary constraints, and what simplifications are acceptable—the model performs this structural transformation reliably. The expert verifies that the simplification does not introduce distortion; the AI provides speed and variation.
Structure and outline generation. Technical reports, white papers, and research proposals follow established genre conventions for section organization, information hierarchy, and scope framing. AI can generate compliant structural outlines that save writers the overhead of genre scaffolding—but the content of each section must be human-authored from verified sources.
Consistency checking across a long document. Asking AI to review a long technical document for terminological consistency—"flag every instance where the term 'user' and 'customer' are used interchangeably, and list all instances of each"—leverages the model's strong pattern-matching on provided text without requiring it to generate factual claims from training data.
Boilerplate and standard section drafting. Many technical documents contain sections that are structurally identical across projects—standard disclaimers, methodology descriptions for established procedures, regulatory compliance language. AI can draft these sections efficiently, with human review against the applicable standards or regulations.
The use of AI in academic writing has generated substantial institutional policy development since 2022. The landscape is complex because policies vary widely by institution, department, assignment type, and even individual instructor—and they are changing rapidly.
By 2024, the majority of major research universities in the United States, United Kingdom, and Australia had adopted formal AI use policies. Most policies distinguish between three categories of use: prohibited use (AI-generated content submitted as the student's own work), permitted use with disclosure (AI used as a tool in a disclosed workflow), and pedagogically integrated use (AI explicitly incorporated into the assignment design). The critical variable is transparency: most institutional policies do not categorically prohibit AI use but require disclosure and clear human intellectual ownership of the final work.
For researchers, the dominant emerging standard—reflected in policies from Nature, Science, Cell, and most major journal publishers by 2024—is that AI cannot be listed as an author, that any AI use in manuscript preparation must be disclosed in the methods section, and that human authors are fully responsible for the accuracy and integrity of all content regardless of what tools were used. This places the accuracy verification burden squarely on the human researcher, making the source-first discipline from journalism equally applicable in academic contexts.
In technical and academic writing, apply the zero-trust rule to every specific claim in AI-generated text: every number, every citation, every organizational name, every date, every attributed quote must be verified against a primary source before use. The genre's formal conventions are not evidence of accuracy—they are the camouflage.
In this lab you will practice identifying citation confabulation risks, apply the zero-trust verification rule to AI-generated technical content, and explore the legitimate uses of AI in technical and academic writing workflows. At least 3 exchanges to complete the lab.