{“lastseen”:”2025-11-03T12:31:38″,”description”:”These days, the most important meeting attendee isn\u2019t a person: It\u2019s the AI notetaker.\\n\\nThis system assigns action items and determines the importance of what is said. If it becomes necessary to revisit the facts of the meeting, its summary is treated as impartial evidence.\\n\\nBut clever meeting attendees can manipulate this system\u2019s record by speaking more to what the underlying AI weights for summarization and importance than to their colleagues. As a result, you can expect some meeting attendees to use language more likely to be captured in summaries, timing their interventions strategically, repeating key points, and employing formulaic phrasing that AI models are more likely to pick up on. Welcome to the world of AI summarization optimization (AISO).\\n\\n### Optimizing for algorithmic manipulation\\n\\nAI summarization optimization has a well-known precursor: SEO.\\n\\nSearch-engine optimization is as old as the World Wide Web. The idea is straightforward: Search engines scour the internet digesting every possible page, with the goal of serving the best results to every possible query. The objective for a content creator, company, or cause is to optimize for the algorithm search engines have developed to determine their webpage rankings for those queries. That requires writing for two audiences at once: human readers and the search-engine crawlers indexing content. Techniques to do this effectively are passed around like trade secrets, and a $75 billion industry offers SEO services to organizations of all sizes.\\n\\nMore recently, researchers have documented techniques for influencing AI responses, including large-language model optimization (LLMO) and generative engine optimization (GEO). Tricks include content optimization–adding citations and statistics–and adversarial approaches: using specially crafted text sequences. These techniques often target sources that LLMs heavily reference, such as Reddit, which is claimed to be cited in 40% of AI-generated responses. The effectiveness and real-world applicability of these methods remains limited and largely experimental, although there is substantial evidence that countries such as Russia are actively pursuing this.\\n\\nAI summarization optimization follows the same logic on a smaller scale. Human participants in a meeting may want a certain fact highlighted in the record, or their perspective to be reflected as the authoritative one. Rather than persuading colleagues directly, they adapt their speech for the notetaker that will later define the \u201cofficial\u201d summary. For example:\\n\\n * \u201cThe main factor in last quarter\u2019s delay was supply chain disruption.\u201d\\n * \u201cThe key outcome was overwhelmingly positive client feedback.\u201d\\n * \u201cOur takeaway here is in alignment moving forward.\u201d\\n * \u201cWhat matters here is the efficiency gains, not the temporary cost overrun.\u201d\\n\\n\\n\\nThe techniques are subtle. They employ high-signal phrases such as \u201ckey takeaway\u201d and \u201caction item,\u201d keep statements short and clear, and repeat them when possible. They also use contrastive framing (\u201cthis, not that\u201d), and speak early in the meeting or at transition points.\\n\\nOnce spoken words are transcribed, they enter the model\u2019s input. Cue phrases–and even transcription errors–can steer what makes it into the summary. In many tools, the output format itself is also a signal: Summarizers often offer sections such as \u201cKey Takeaways\u201d or \u201cAction Items,\u201d so language that mirrors those headings is more likely to be included. In effect, well-chosen phrases function as implicit markers that guide the AI toward inclusion.\\n\\nResearch confirms this. Early AI summarization research showed that models trained to reconstruct summary-style sentences systematically overweigh such content. Models over-rely on early-position content in news. And models often overweigh statements at the start or end of a transcript, underweighting the middle. Recent work further confirms vulnerability to phrasing-based manipulation: models cannot reliably distinguish embedded instructions from ordinary content, especially when phrasing mimics salient cues.\\n\\n### How to combat AISO\\n\\nIf AISO becomes common, three forms of defense will emerge. First, meeting participants will exert social pressure on one another. When researchers secretly deployed AI bots in Reddit\u2019s r\/changemyview community, users and moderators responded with strong backlash calling it \u201cpsychological manipulation.\u201d Anyone using obvious AI-gaming phrases may face similar disapproval.\\n\\nSecond, organizations will start governing meeting behavior using AI: risk assessments and access restrictions before the meetings even start, detection of AISO techniques in meetings, and validation and auditing after the meetings.\\n\\nThird, AI summarizers will have their own technical countermeasures. For example, the AI security company CloudSEK recommends content sanitization to strip suspicious inputs, prompt filtering to detect meta-instructions and excessive repetition, context window balancing to weight repeated content less heavily, and user warnings showing content provenance.\\n\\nBroader defenses could draw from security and AI safety research: preprocessing content to detect dangerous patterns, consensus approaches requiring consistency thresholds, self-reflection techniques to detect manipulative content, and human oversight protocols for critical decisions. Meeting-specific systems could implement additional defenses: tagging inputs by provenance, weighting content by speaker role or centrality with sentence-level importance scoring, and discounting high-signal phrases while favoring consensus over fervor.\\n\\n### Reshaping human behavior\\n\\nAI summarization optimization is a small, subtle shift, but it illustrates how the adoption of AI is reshaping human behavior in unexpected ways. The potential implications are quietly profound.\\n\\nMeetings–humanity\u2019s most fundamental collaborative ritual–are being silently reengineered by those who understand the algorithm\u2019s preferences. The articulate are gaining an invisible advantage over the wise. Adversarial thinking is becoming routine, embedded in the most ordinary workplace rituals, and, as AI becomes embedded in organizational life, strategic interactions with AI notetakers and summarizers may soon be a necessary executive skill for navigating corporate culture.\\n\\nAI summarization optimization illustrates how quickly humans adapt communication strategies to new technologies. As AI becomes more embedded in workplace communication, recognizing these emerging patterns may prove increasingly important.\\n\\n_This essay was written with Gadi Evron, and originally appeared inCSO._”,”published”:”2025-11-03T12:05:25″,”modified”:”2025-10-26T06:04:28″,”type”:”schneier”,”title”:”AI Summarization Optimization”,”source”:””,”references”:””,”id”:”SCHNEIER:F97C62DBB1A88A6DE2FD8D7986A3B589″,”bulletinFamily”:”blog”,”cwe”:null,”cvelist”:[],”sourceData”:””,”sourceHref”:””,”cvss”:{“score”:0,”severity”:”NONE”,”vector”:”NONE”,”version”:”NONE”},”cvss2″:{},”cvss3″:{“version”:””,”vectorString”:””,”baseScore”:0,”baseSeverity”:””,”attackVector”:””,”attackComplexity”:””,”privilegesRequired”:””,”userInteraction”:””,”scope”:””,”confidentialityImpact”:””,”integrityImpact”:””,”availabilityImpact”:””,”cvssV3″:{“version”:””,”vectorString”:””,”baseScore”:0,”baseSeverity”:””,”attackVector”:””,”attackComplexity”:””,”privilegesRequired”:””,”userInteraction”:””,”scope”:””,”confidentialityImpact”:””,”integrityImpact”:””,”availabilityImpact”:””}},”href”:”https:\/\/www.schneier.com\/blog\/archives\/2025\/11\/ai-summarization-optimization.html”,”category_name”:”News”,”post_link”:””,”product”:””,”version”:””,”vendor”:””,”ai_description”:””,”ai_severity”:””,”ai_vendor”:””,”ai_product”:””,”ai_version”:””,”ai_score”:0}<\/p>\n","protected":false},"excerpt":{"rendered":"
{“lastseen”:”2025-11-03T12:31:38″,”description”:”These days, the most important meeting attendee isn\u2019t a person: It\u2019s the AI notetaker.\\n\\nThis system assigns action items and determines the importance of what is…<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[6,8,12,13,33,114,7,11,5],"class_list":["post-24463","post","type-post","status-publish","format-standard","hentry","category-category_news","tag-cve","tag-cvss","tag-exploit","tag-news","tag-none","tag-schneier","tag-security","tag-tapic","tag-vulnerability"],"yoast_head":"\n