AI Destroys Institutions by Eroding Expertise—Voice AI for Demos Proves Why Reading DOM Preserves Knowledge, While Generation Destroys Trust

# AI Destroys Institutions by Eroding Expertise—Voice AI for Demos Proves Why Reading DOM Preserves Knowledge, While Generation Destroys Trust *Hacker News #1 (81 points, 44 comments, 43min): Stanford professors publish "How AI Destroys Institutions"—AI systems erode expertise, short-circuit decision-making, and isolate people. "Current AI systems are a death sentence for civic institutions." This applies to product demos: Chatbots erode user expertise, short-circuit exploration, and isolate users from actual product knowledge.* --- ## The Academic Verdict: AI Is Killing Institutions Stanford Law School's Center for Internet and Society just published a paper with a brutal thesis: **"Current AI systems are a death sentence for civic institutions, and we should treat them as such."** Authors Woodrow Hartzog (Boston University School of Law) and Jessica M. Silbey (Boston University School of Law) make one simple point in their essay titled "How AI Destroys Institutions": "AI systems are built to function in ways that degrade and are likely to destroy our crucial civic institutions. The affordances of AI systems have the effect of **eroding expertise, short-circuiting decision-making, and isolating people from each other**." The paper focuses on civic institutions—universities, the rule of law, a free press—but the pattern applies to product demos too. Chatbot demos erode user expertise, short-circuit product exploration, and isolate users from actual product knowledge. ## The Three Ways AI Destroys Institutions From the abstract: "Civic institutions—the rule of law, universities, and a free press—are the backbone of democratic life. They are the mechanisms through which complex societies encourage cooperation and stability, while also adapting to changing circumstances. **The real superpower of institutions is their ability to evolve and adapt within a hierarchy of authority and a framework for roles and rules while maintaining legitimacy in the knowledge produced and the actions taken.**" The paper identifies three mechanisms of institutional destruction: ### Mechanism #1: Eroding Expertise "The affordances of AI systems have the effect of **eroding expertise**." **What this means for institutions:** - Universities produce expert knowledge through research, peer review, and credentialing - AI systems bypass this process—generating text that looks authoritative without expertise - Users can't distinguish expert knowledge from AI-generated content - Trust in institutions that certify expertise collapses **What this means for demos:** - Product teams build expertise about features, workflows, edge cases - Chatbot demos bypass this expertise—generating answers from training data - Users can't distinguish actual product behavior from hallucinated features - Trust in product documentation collapses ### Mechanism #2: Short-Circuiting Decision-Making "The affordances of AI systems have the effect of **short-circuiting decision-making**." **What this means for institutions:** - Democratic decision-making requires deliberation, debate, consensus-building - AI systems provide instant answers—bypassing the process of working through complexity - Institutions lose the ability to evolve through collective reasoning - Decisions are made without the legitimacy that comes from process **What this means for demos:** - Product understanding requires exploration, experimentation, discovery - Chatbot demos provide instant answers—bypassing the process of learning the product - Users lose the ability to build mental models through hands-on experience - Product adoption happens without the confidence that comes from direct knowledge ### Mechanism #3: Isolating People from Each Other "The affordances of AI systems have the effect of **isolating people from each other**." **What this means for institutions:** - Institutions function through interpersonal relationships, shared commitment, collective goals - AI systems mediate human interaction—replacing person-to-person exchange with person-to-AI interface - "This happens through the machinations of interpersonal relationships within those institutions, which broaden perspectives and strengthen shared commitment to civic goals." - When AI replaces human interaction, institutions lose their social fabric **What this means for demos:** - Product mastery happens through user-to-interface exploration, discovering features, understanding workflows - Chatbot demos mediate this exploration—replacing user-to-product interaction with user-to-AI conversation - Users should explore the actual product interface to build understanding - When chatbots replace direct exploration, users lose connection to the product itself ## Why Institutions Have "Superpower of Evolution" The paper emphasizes what makes institutions valuable: "The real superpower of institutions is their ability to **evolve and adapt** within a hierarchy of authority and a framework for roles and rules while **maintaining legitimacy in the knowledge produced and the actions taken**." **Two key capabilities:** ### Capability #1: Evolution Through Process Institutions don't just produce knowledge—they **evolve their knowledge production methods** through: - Peer review (academia) - Precedent and argument (law) - Editorial standards and fact-checking (journalism) **The process creates legitimacy.** A peer-reviewed paper is trustworthy not because of the specific findings, but because it survived a rigorous vetting process. **AI destroys this:** Generated text bypasses the process. There's no peer review, no editorial oversight, no precedent analysis. The output looks authoritative but lacks the legitimacy that process provides. ### Capability #2: Transparency and Accountability "Purpose-driven institutions built around **transparency, cooperation, and accountability** empower individuals to take intellectual risks and challenge the status quo." **Why transparency matters:** - Academic research: Methods are documented, data is shared, experiments are replicable - Legal system: Court proceedings are public, precedents are cited, decisions are explained - Free press: Sources are disclosed, corrections are published, editorial process is visible **Why AI destroys this:** - Training data is opaque - Generation process is black box - Corrections don't propagate (hallucination persists across conversations) - No accountability (who's responsible for wrong answer?) ## The Pattern: Institutions Need Legitimacy, AI Provides None The Stanford paper's core insight: **Institutions derive power from legitimacy. Legitimacy comes from process, transparency, and interpersonal relationships. AI systems bypass all three.** ### How Universities Create Legitimacy **Process:** 1. Researcher conducts study 2. Submits to peer-reviewed journal 3. Experts review methodology 4. Paper accepted or rejected based on rigor 5. Published with citation of sources 6. Other researchers can replicate **Result:** Trust in the knowledge produced because the process is rigorous. ### How AI Destroys Legitimacy **Process:** 1. User asks question 2. LLM generates answer from training data 3. No review of accuracy 4. Answer displayed as authoritative 5. No citations of sources 6. No way to verify or replicate **Result:** No trust in the knowledge produced because there's no process. ## Why Chatbot Demos Follow the Same Pattern The Stanford paper focuses on civic institutions, but the destruction pattern applies to product demos: ### Demo Pattern #1: Eroding Expertise (Product Knowledge) **Product team expertise:** - Engineers know how features work - Designers know intended workflows - Support team knows edge cases - Documentation writers know common questions **How chatbot demos erode this:** - Generate answers from training data (not actual product code) - May hallucinate features that don't exist - Can't distinguish between "product works this way" and "training data says this" - Users learn incorrect information **Voice AI preserves expertise:** - Reads actual DOM structure (not generating from training data) - Describes what actually exists on page - Can't hallucinate features (only reads what's there) - Users learn correct information ### Demo Pattern #2: Short-Circuiting Decision-Making (Product Exploration) **Product mastery through exploration:** 1. User opens dashboard 2. Sees available features 3. Clicks through sections 4. Discovers workflows 5. Builds mental model through direct experience 6. Confident they understand how it works **How chatbot demos short-circuit this:** - User asks "How does analytics work?" - Chatbot generates description from training data - User reads explanation but doesn't explore interface - No hands-on experience, no mental model building - Uncertain if description matches actual product **Voice AI preserves exploration:** - User opens analytics page - Voice AI: "This page shows three sections: [reads actual headings]" - User explores each section with contextual guidance - Hands-on experience builds mental model - Confident because they experienced the product directly ### Demo Pattern #3: Isolating People from Product (User-Interface Relationship) **Product adoption through direct interaction:** - User interacts with actual interface - Discovers affordances (buttons, menus, settings) - Understands spatial relationships (where features live) - Builds muscle memory for workflows - Feels connected to the product **How chatbot demos isolate users:** - User asks questions to chatbot instead of exploring product - Chatbot describes features instead of user discovering them - Conversation replaces exploration - User knows *about* product but hasn't *used* product - Feels disconnected from actual interface **Voice AI preserves connection:** - User explores actual interface - Voice AI provides contextual guidance while user navigates - Exploration augmented, not replaced - User discovers features through hands-on interaction - Feels connected to product through direct experience ## The "Death Sentence" Quote Explains Why Chatbots Fail The paper's conclusion: "Current AI systems are a death sentence for civic institutions, and we should treat them as such." **Why this language matters:** Not "AI is concerning" or "AI poses challenges." The language is **"death sentence."** **Why such strong language?** Because the destruction isn't a side effect—it's architectural. AI systems are **built in ways that inherently erode** the features that make institutions work. **The three architectural incompatibilities:** ### Architectural Incompatibility #1: Generation vs. Expertise **Institutions:** Knowledge is produced through expert-validated processes **AI systems:** Content is generated from pattern-matching in training data **These are fundamentally incompatible.** You can't have both expertise-validation and autoregressive generation. ### Architectural Incompatibility #2: Instant Answers vs. Deliberation **Institutions:** Decisions emerge through deliberation, debate, consensus-building **AI systems:** Answers are produced instantly through next-token prediction **These are fundamentally incompatible.** You can't have both deliberative process and immediate generation. ### Architectural Incompatibility #3: Mediated Interaction vs. Direct Relationships **Institutions:** Value emerges through interpersonal relationships and shared commitment **AI systems:** Interaction is mediated through person-to-AI interface **These are fundamentally incompatible.** You can't have both direct human relationships and AI-mediated conversations. ## Why Voice AI Avoids All Three Incompatibilities The Stanford paper identifies why **generation-based AI** destroys institutions. Voice AI for demos avoids this by **reading instead of generating**. ### Voice AI Solution #1: Reading Preserves Expertise **Instead of generating from training data:** Voice AI reads actual DOM structure **Why this preserves expertise:** - Product team's knowledge is embedded in the page structure - Headings, buttons, sections reflect actual product design - Voice AI describes what developers built, not what LLM imagines - User learns from actual product, not hallucinated version **The pattern:** Reading transmits expertise. Generation erodes it. ### Voice AI Solution #2: Guidance Enables Exploration **Instead of instant answers that bypass exploration:** Voice AI provides contextual guidance during exploration **Why this enables deliberation:** - User must still explore interface themselves - Voice AI augments understanding, doesn't replace experience - Mental model building happens through hands-on interaction - User builds confidence through direct discovery **The pattern:** Guidance enables process. Generation short-circuits it. ### Voice AI Solution #3: Augmentation Preserves Connection **Instead of mediated conversation replacing direct interaction:** Voice AI augments direct product interaction **Why this preserves connection:** - User interacts with actual product interface - Voice AI explains what user is looking at - Exploration happens in product, not in chat window - User builds relationship with product, not with AI **The pattern:** Augmentation preserves connection. Mediation isolates. ## The Three Institutional Failures That Parallel Demo Failures ### Institutional Failure #1: Academic Credibility Collapse **Scenario:** Student submits AI-generated essay **Process breakdown:** 1. No original research conducted 2. No sources verified 3. No expertise developed 4. Essay looks authoritative but is meaningless 5. Can't distinguish from expert-written work **Result:** Trust in academic credentials collapses. If essays are AI-generated, what does an A grade certify? ### Demo Failure #1: Product Understanding Collapse **Scenario:** User asks chatbot demo "How does this feature work?" **Process breakdown:** 1. No product exploration conducted 2. No interface examined 3. No hands-on experience gained 4. Answer looks authoritative but may be hallucinated 5. Can't distinguish from actual product behavior **Result:** Trust in product knowledge collapses. If demos are chatbot-generated, what does "I understand the product" mean? ### Institutional Failure #2: Legal Precedent Bypass **Scenario:** Judge uses AI to draft legal opinion **Process breakdown:** 1. No case law research conducted 2. No precedent analysis performed 3. No legal reasoning documented 4. Opinion looks authoritative but cites fake cases 5. Can't distinguish from expert legal analysis **Result:** Trust in legal system collapses. If opinions are AI-generated, what legitimacy do court decisions have? ### Demo Failure #2: Feature Discovery Bypass **Scenario:** User asks chatbot "What can this product do?" **Process breakdown:** 1. No feature exploration conducted 2. No workflow experimentation performed 3. No capability testing documented 4. Answer looks comprehensive but may omit features 5. Can't distinguish from exhaustive product knowledge **Result:** Trust in product evaluation collapses. If demos are chatbot-described, how do users know what product actually does? ### Institutional Failure #3: Journalistic Source Verification Lost **Scenario:** News org publishes AI-generated article **Process breakdown:** 1. No sources interviewed 2. No facts verified 3. No editorial oversight applied 4. Article looks reported but is fabricated 5. Can't distinguish from actual journalism **Result:** Trust in press collapses. If articles are AI-generated, what does "published in [outlet]" certify? ### Demo Failure #3: Product Interface Understanding Lost **Scenario:** User relies on chatbot to learn product without exploring interface **Process breakdown:** 1. No interface navigation performed 2. No button/menu discovery completed 3. No workflow experimentation done 4. Knowledge looks complete but is abstract 5. Can't distinguish from hands-on product mastery **Result:** Trust in product adoption collapses. If demos are chatbot-explained, does user actually know how to use product? ## The "Legitimacy in Knowledge Produced" Standard The Stanford paper emphasizes: "maintaining legitimacy in the knowledge produced and the actions taken." **What creates legitimacy:** ### In Academic Institutions - Peer review process - Replication of experiments - Citation of sources - Transparency of methods - Expert evaluation ### In Legal Institutions - Adversarial process - Precedent analysis - Public proceedings - Documented reasoning - Appellate review ### In Journalistic Institutions - Source verification - Fact-checking - Editorial oversight - Corrections policy - Byline accountability **What AI destroys:** All of the above. There's no peer review of generated text, no precedent analysis of legal opinions, no source verification of articles. ### In Product Demos **What creates legitimacy in product knowledge:** - Direct interface exploration - Hands-on feature testing - Workflow experimentation - Edge case discovery - Personal verification of capabilities **What chatbot demos destroy:** All of the above. There's no interface exploration of generated answers, no hands-on testing of described features, no workflow experimentation in explained processes. **What voice AI preserves:** All of the above. Voice AI guides user through actual interface exploration, hands-on feature testing, workflow experimentation. ## Why "Eroding Expertise" Is the Core Problem Of the three mechanisms (eroding expertise, short-circuiting decision-making, isolating people), **eroding expertise** is the foundation. **Why expertise comes first:** Without expertise, you can't make good decisions (even with deliberation). Without expertise, you can't build strong relationships (no shared knowledge foundation). **Expertise is the substrate on which institutions are built.** ### How Universities Build Expertise **Process:** 1. Years of study in field 2. Original research contribution 3. Peer validation of work 4. Credentialing (PhD, professor, tenure) 5. Continued research and publication 6. Recognition by expert community **Result:** When someone with these credentials speaks, you trust their expertise. ### How AI Erodes Expertise **Process:** 1. User prompts LLM 2. LLM generates authoritative-sounding text 3. No validation of accuracy 4. Text presented as knowledge 5. No way to verify expertise 6. No accountability for errors **Result:** When AI generates text, you can't trust its expertise—but it looks identical to expert knowledge. **The crisis:** Users can't distinguish expert knowledge from AI-generated content. So all knowledge becomes suspect. ### How Chatbot Demos Erode Product Expertise **Process:** 1. User asks question 2. Chatbot generates answer from training data 3. No validation against actual product 4. Answer presented as product knowledge 5. No way to verify accuracy 6. No accountability for hallucinations **Result:** When chatbot describes product, you can't trust its knowledge—but it looks identical to expert product documentation. **The crisis:** Users can't distinguish actual product behavior from hallucinated features. So all product knowledge becomes suspect. ### How Voice AI Preserves Product Expertise **Process:** 1. User explores page 2. Voice AI reads actual DOM structure 3. Describes what exists (not generating) 4. User sees what Voice AI describes 5. Direct verification (user can see it) 6. Impossible to hallucinate (reading, not generating) **Result:** When Voice AI describes product, you can trust it's accurate—because it's reading what's actually there. **The solution:** Users verify Voice AI guidance by looking at page. Expertise is preserved through direct observation. ## The "Interpersonal Relationships" Insight The Stanford paper emphasizes: "This happens through the machinations of interpersonal relationships within those institutions, which broaden perspectives and strengthen shared commitment to civic goals." **Why interpersonal relationships matter:** Institutions aren't just processes—they're **people working together** within those processes. ### How Universities Function Through Relationships **Not just:** Submit paper → Get reviewed → Publish **Actually:** - Discuss ideas with colleagues - Present at conferences - Receive feedback - Debate interpretations - Build on each other's work - Form research communities **The value:** Perspectives are broadened, commitments are strengthened, knowledge is refined through social interaction. ### How AI Isolates People **AI replaces interpersonal interaction:** - Instead of: Discuss with colleague → Get perspective → Refine thinking - AI provides: Ask chatbot → Get answer → Accept or reject **No broadening of perspectives.** Just one-to-one human-AI interaction. **No strengthening of commitment.** Just transactional question-answer exchange. ### How Chatbot Demos Isolate Users from Product **Chatbot replaces product interaction:** - Instead of: Explore interface → Discover feature → Understand workflow - Chatbot provides: Ask question → Get answer → Move on **No building of mental models.** Just abstract descriptions without hands-on experience. **No strengthening of product mastery.** Just transactional information exchange. ### How Voice AI Preserves Product Interaction **Voice AI augments product interaction:** - Explore interface → Voice AI explains what you're seeing → Discover feature → Understand workflow **Mental models build through direct experience** guided by contextual explanations. **Product mastery strengthens through hands-on exploration** augmented by accurate descriptions. ## Why "Anathema to Evolution" Matters The Stanford paper: "These systems are anathema to the kind of evolution, transparency, cooperation, and accountability that give vital institutions their purpose and sustainability." **"Anathema"** is strong language. It means fundamentally opposed, incompatible at a core level. **Why AI is anathema to institutional evolution:** ### Institutions Evolve Through Feedback Loops **Academic evolution:** - Research → Publication → Criticism → Improved research → New publication - Methods are refined through community critique - Knowledge accumulates through transparent correction **Legal evolution:** - Case → Decision → Appeal → Precedent → Future cases - Law adapts through adversarial challenge - Justice improves through documented reasoning **Journalistic evolution:** - Story → Publication → Fact-check → Correction → Better standards - Accuracy improves through accountability - Trust rebuilds through transparency **The pattern:** Mistakes are visible, corrections are documented, improvements are systematic. ### AI Cannot Evolve Through Feedback **AI generation:** - Prompt → Generation → User reads → Next prompt - Mistakes aren't visible (hallucinations look authoritative) - Corrections don't propagate (each generation is independent) - Improvements aren't systematic (same prompt can hallucinate differently) **The problem:** There's no evolutionary pressure. Bad outputs don't create systematic improvements. ### Chatbot Demos Cannot Evolve User Understanding **Chatbot demo interaction:** - Question → Generated answer → User accepts or rejects → Next question - Mistakes aren't visible (hallucinated features look real) - Corrections don't propagate (each answer is independent) - Understanding doesn't accumulate (no hands-on experience to build on) **The problem:** User can't build on previous knowledge because there's no direct experience foundation. ### Voice AI Enables Evolutionary Understanding **Voice AI interaction:** - Explore page → Voice AI describes structure → User discovers feature → Deeper exploration - Mistakes are visible (user sees what Voice AI describes) - Understanding propagates (hands-on experience builds mental model) - Knowledge accumulates (each exploration builds on previous discoveries) **The solution:** User's understanding evolves through guided exploration of actual product. ## The Three Questions the Stanford Paper Raises for Demos ### Question #1: Does Your Demo Erode Expertise or Preserve It? **Eroding expertise:** - Generating answers from training data - Describing features that may not exist - Explaining workflows user never experiences - Creating illusion of knowledge without hands-on understanding **Preserving expertise:** - Reading actual page structure - Describing only what exists - Guiding user through workflows they directly experience - Building knowledge through hands-on exploration **Voice AI choice:** Preserve expertise by reading DOM, not generating descriptions. ### Question #2: Does Your Demo Short-Circuit Exploration or Enable It? **Short-circuiting exploration:** - Providing instant answers that bypass hands-on discovery - Explaining product without user experiencing it - Creating abstract knowledge divorced from interface interaction - Replacing exploration with conversation **Enabling exploration:** - Providing contextual guidance during hands-on discovery - Explaining what user is currently experiencing - Building concrete knowledge through interface interaction - Augmenting exploration with explanations **Voice AI choice:** Enable exploration by providing guidance, not replacement. ### Question #3: Does Your Demo Isolate Users from Product or Connect Them? **Isolating users:** - Replacing product interaction with chatbot conversation - Describing interface instead of user exploring it - Creating knowledge "about product" without "using product" - Mediating relationship between user and interface **Connecting users:** - Augmenting product interaction with voice guidance - Explaining interface while user explores it - Creating knowledge through direct product use - Preserving relationship between user and interface **Voice AI choice:** Connect users by augmenting, not mediating. ## Why "Death Sentence" Isn't Hyperbole The Stanford paper uses extreme language: "AI systems are a death sentence for civic institutions." **Is this hyperbole?** No. Here's why: ### Death Sentence = Inevitable Destruction **Not:** "AI might damage institutions" **But:** "AI will destroy institutions" **The certainty comes from architectural incompatibility:** - Institutions require expertise → AI erodes expertise - Institutions require deliberation → AI short-circuits deliberation - Institutions require interpersonal relationships → AI isolates people **You can't fix this.** It's not a bug you can patch. It's fundamental to how generative AI works. ### For Civic Institutions **Universities:** If AI-generated essays are indistinguishable from student work, what does a degree certify? **Legal system:** If AI-generated opinions cite fake cases, what legitimacy do decisions have? **Free press:** If AI-generated articles fabricate sources, what does publication mean? **These aren't hypothetical.** These failures are already happening. ### For Product Demos **If chatbot demos hallucinate features:** What does "I tried the demo" mean? **If generated answers contradict actual product:** What does "I understand the product" mean? **If users learn from conversations, not interface:** What does "I'm ready to use it" mean? **These aren't hypothetical.** These failures are already happening in SaaS demos. ## The Alternative: Voice AI as Institutional Preservation The Stanford paper identifies the problem: Generative AI destroys institutions by eroding expertise, short-circuiting process, and isolating people. Voice AI for demos provides the solution: **Read instead of generate.** ### Why Reading Preserves Institutional Values **Expertise preservation:** Voice AI reads page structure created by product team (transmits expertise) **Process preservation:** Voice AI guides exploration, doesn't replace it (enables deliberation) **Relationship preservation:** Voice AI augments product interaction, doesn't mediate it (maintains connection) **The pattern:** Reading transmits knowledge from source (product team) to user directly. Generation creates new content divorced from source expertise. ### The Parallel to Academic Citation **Bad academic practice:** Rewrite someone else's work in your own words without citation **Why it's bad:** Expertise attribution is lost, original source is obscured, knowledge provenance is hidden **Good academic practice:** Quote directly with citation to original source **Why it's good:** Expertise is attributed, original source is accessible, knowledge provenance is clear **Chatbot demos:** Paraphrase training data about product in generated text **Why it's bad:** Product team expertise is lost, actual product behavior is obscured, knowledge source is hidden **Voice AI:** "Quote" page structure by reading DOM directly **Why it's good:** Product team expertise is attributed (in page structure), actual product is accessible (user sees it), knowledge source is clear (reading, not generating) ## The Verdict: Generation Destroys, Reading Preserves The Stanford paper's conclusion applies to product demos: **Current AI systems are a death sentence for civic institutions** because they erode expertise, short-circuit decision-making, and isolate people. **Current chatbot demos are a death sentence for product understanding** because they erode product expertise (hallucinations), short-circuit exploration (instant answers), and isolate users from product (mediated conversation). **Voice AI avoids this death sentence** by reading page structure (preserves expertise), guiding exploration (enables process), and augmenting interaction (maintains connection). ## The Three Institutional Principles Voice AI Follows ### Institutional Principle #1: Expertise Through Process **Universities:** Knowledge gains legitimacy through peer review process **Voice AI:** Product knowledge gains legitimacy through DOM reading (not generation) **Why process matters:** It creates trust. You trust peer-reviewed research because it survived scrutiny. You trust Voice AI descriptions because they're reading actual page structure. ### Institutional Principle #2: Evolution Through Transparency **Legal system:** Precedent evolves through documented decisions and public reasoning **Voice AI:** Product understanding evolves through visible exploration guided by transparent descriptions **Why transparency matters:** It enables improvement. You can critique legal reasoning because it's documented. You can verify Voice AI guidance because you see what it describes. ### Institutional Principle #3: Legitimacy Through Relationships **Free press:** Journalism gains legitimacy through reporter-source relationships and editorial collaboration **Voice AI:** Product knowledge gains legitimacy through user-interface relationship and exploration experience **Why relationships matter:** They create accountability. You trust journalism because reporters stake reputation on accuracy. You trust product knowledge because you experienced interface directly. ## The Pattern: AI That Reads Builds Institutions, AI That Generates Destroys Them The Stanford paper identifies **generative AI** as institutional poison. The solution isn't "no AI in institutions." The solution is **AI that reads instead of generates.** ### AI That Reads in Academia **Not:** LLM writes research paper **But:** AI reads existing papers, helps researcher find relevant sources, highlights methodology patterns **Preserves expertise:** Research still conducted by human expert **Enables process:** AI augments literature review, doesn't replace research **Maintains relationships:** Researcher collaborates with AI tool, still engages with academic community ### AI That Reads in Law **Not:** LLM drafts legal opinion **But:** AI reads case law, helps lawyer find precedents, highlights relevant statutes **Preserves expertise:** Legal reasoning still performed by human lawyer **Enables process:** AI augments research, doesn't replace analysis **Maintains relationships:** Lawyer uses AI tool, still engages with court and opposing counsel ### AI That Reads in Journalism **Not:** LLM writes news article **But:** AI reads documents, helps journalist verify facts, highlights contradictions in sources **Preserves expertise:** Reporting still performed by human journalist **Enables process:** AI augments fact-checking, doesn't replace investigation **Maintains relationships:** Journalist uses AI tool, still engages with sources and editors ### AI That Reads in Product Demos **Not:** Chatbot generates product descriptions **But:** Voice AI reads page structure, helps user navigate interface, highlights available features **Preserves expertise:** Product built by human team, Voice AI reads their work **Enables process:** AI augments exploration, doesn't replace hands-on experience **Maintains relationships:** User uses AI tool, still engages with actual product interface ## The Alternative: Treating Generation as the Problem, Not AI Itself The Stanford paper warns: "we should treat them as such" (death sentence). **This doesn't mean "ban AI."** This means **recognize the specific affordances that cause destruction** and avoid them. ### The Destructive Affordances **From the paper:** "The affordances of AI systems have the effect of eroding expertise, short-circuiting decision-making, and isolating people from each other." **Affordances = the properties of a system that enable specific uses** **Which affordances destroy institutions?** 1. **Generation** (creating content from training data) → Erodes expertise 2. **Instant answers** (bypassing deliberation) → Short-circuits decision-making 3. **Conversation mediation** (replacing human interaction) → Isolates people **Which affordances preserve institutions?** 1. **Reading** (extracting information from source) → Transmits expertise 2. **Contextual guidance** (augmenting process) → Enables decision-making 3. **Augmentation** (enhancing human interaction) → Connects people ### Voice AI as Institutional Preservation Through Affordance Design **Voice AI avoids destructive affordances:** - Reads DOM instead of generating from training data - Provides guidance during exploration instead of instant replacement answers - Augments product interaction instead of mediating it through conversation **Voice AI implements preserving affordances:** - Reading transmits product team expertise to user - Guidance enables user's exploration process - Augmentation maintains user-product connection **The lesson:** AI isn't inherently destructive. Specific affordances (generation, mediation, instant answers) are destructive. Different affordances (reading, augmentation, guidance) are preserving. --- *Demogod's voice AI reads your site's DOM directly—preserving product expertise through reading (not generation), enabling user exploration through guidance (not replacement), and maintaining user-product connection through augmentation (not mediation). Like academic citation vs. plagiarism: reading with attribution preserves institutional legitimacy. One line of code. Zero generation. [Try it on your site](https://demogod.me).*