A workplace AI summary is only as trustworthy as its resistance to pressure—whether that's misleading formatting, loaded questions, or incomplete source material. Building robustness means testing your AI system against realistic distortions and checking that it maintains accuracy when pushed.
Adversarial robustness refers to how consistently an AI model produces reliable outputs when faced with inputs designed to manipulate or confuse it. In workplace contexts, you're not concerned with hostile actors testing your personal AI deployment (low likelihood). Instead, you're concerned with whether your AI tools reliably extract accurate information and summaries even when source materials contain contradictory information, bias, or intentional obfuscation.
Consider an employee handbook with contradictory policies. Section 3 says "remote work requires manager approval," but Section 8 says "all approved roles may work remotely." These aren't adversarial attacks—they're realistic workplace documentation. An adversarially non-robust AI might confidently summarize either interpretation depending on which it encountered first, or it might average them into a summary that's technically incorrect.
These are related but distinct problems. Hallucination means the AI invents information. Lack of robustness means the AI's output varies unreliably based on input ordering, formatting, or minor wording changes. You ask Claude to summarize three conflicting manager comments about your performance, and depending on how you paste them (comment A first or comment C first), you get different summaries. Neither is hallucinated—both are derived from provided text—but the variation reveals non-robust processing.
Adversarial robustness matters for workplace documentation because you need your summaries to be reproducible and defensible. If you present a summary to HR and they say "that's not what those documents say," you can't explain that "the AI just interpreted it differently this time." Your summary should be the same regardless of formatting or input ordering.
The simplest test: reorder your inputs. Ask Claude to summarize your performance reviews in chronological order. Generate the summary. Then paste the same reviews in reverse chronological order and regenerate. The summaries should be substantively identical. If they contradict each other or emphasize different points, you've identified a robustness issue.
Second test: formatting variation. Paste documents formatted one way (raw text), then paste them formatted another way (bullet points), with identical content. Inconsistent outputs reveal formatting dependency, a robustness weakness.
Third test: summary consistency. Ask the AI to summarize the same material twice in the same conversation. Identical or near-identical outputs indicate robustness. Divergent summaries indicate the AI is sensitive to small perturbations or randomness despite temperature being low.
If you're building documentation to protect yourself from workplace retaliation or gaslighting, robustness is non-negotiable. You might present your AI-generated summary of manager communications to HR. They'll scrutinize it. If they ask Claude the same question independently and get different results, your credibility collapses. "That's just how the AI interpreted it" is not a defense when your entire documentation relies on AI accuracy.
This is why system-of-records approaches matter more than single-shot AI outputs. You're not using AI to generate a summary once and filing it away. You're using AI as a tool to help you create a summary that you then verify independently and file as your own analysis.
Specify the summary format explicitly. Instead of "summarize these performance reviews," ask: "List each review's rating in column 1, feedback themes in column 2, and specific examples in column 3. Present in chronological order." Structured output requests are more robust than open-ended ones.
Break source materials into smaller, explicit chunks. Instead of pasting 10 emails and asking for overall themes, ask the AI to extract key points from email #1, then #2, then #3, then synthesize. This chunked approach reduces sensitivity to ordering or formatting of the full batch.
Request confidence levels. Ask the AI to mark statements as "directly supported by source material," "inferred from context," or "contradicted by other materials." This transparency reveals where robustness might be failing—inferred statements are less robust.
A robust AI should flag when source materials contradict each other. If you ask Claude to summarize policy documents and they contain contradictory provisions, the AI should say "Section 3 states X, while Section 8 states Y" rather than trying to synthesize a single coherent statement. This transparency reveals robustness limits and prevents false confidence in summary accuracy.
Many AI tools default to synthesis rather than contradiction flagging. Override this by explicitly requesting: "If source materials contradict each other, state both positions clearly rather than trying to reconcile them."
Try this: Collect three workplace documents related to a policy or decision (emails, handbook sections, meeting notes—anything with information density). Ask Claude to extract key facts and generate a summary. Copy that prompt exactly and regenerate. Then paste the source documents in different order and regenerate again. Compare all three summaries. Identical outputs indicate robustness. Variations reveal ordering sensitivity. Document these findings as your baseline robustness profile for Claude.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.