Roundtable: How We Really Use AI in Testing  🔗 **Source:** [Roundtable: How We Really Use AI in Testing](https://nicolevanderhoeven.com/blog/20261008-ai-testing-roundtable-hustef/) > [!note]+ Document Notes > tags: [ai-in-testing, software-testing, hustef, ai-ethics, conference-recap] > author: Nicole van der Hoeven > source: nicolevanderhoeven.com > reference: https://nicolevanderhoeven.com/blog/20261008-ai-testing-roundtable-hustef/ > date created: 2026-10-12 11:04 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Roundtable: How We Really Use AI in Testing (HUSTEF 2026) > <!-- deep-read:sections start --> ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--roundtable-how-we-really-use-ai-in-testing-hustef-2026) ## 📝 Summary Nicole van der Hoeven's blog (nicolevanderhoeven.com) published this first-person recap on 2026-10-08 (date taken from the URL). It is a personal blog post: practitioner experience and opinion, not research. It reports one panelist's answers at a HUSTEF 2026 roundtable in Budapest. The other four participants (Marta Firlej, moderating, plus Kristoffer Nordström, Callum Akehurst-Ryan, and Maryia Tuleika) are named, but the post reports none of their answers. - **What happened.** On the final day of HUSTEF 2026, five people held a roundtable on AI in testing. Firlej had run the same "AI open space" exercise with the Polish testing community and brought the crowdsourced uses to the panel (P1-P2). - **The claim.** The author uses AI for two work tasks (reading codebases and test code, and filling quality-engineer gaps on teams) and for private experiments on a local open-weight model (P3). The author keeps current through coworkers and a scheduled digest, not release news (P4), advises against over-trusting AI (P5), refuses to let it write messages to people (P6), and treats ethics as a sliding scale (P7). - **Why it matters.** The post is a concrete inventory of one tester's AI habits, with limits stated. The author also notes the access is "effectively unlimited" for now and expects that to end (P3). - **Key stakeholders.** The author, the four named panelists, the HUSTEF audience, and the Polish testing community whose open-space input seeded the questions. - **Risks and tradeoffs.** The post reports no outcomes, measurements, or error rates. The personal-data experiment (journals, therapy notes, WhatsApp history) trades privacy exposure for insight, mitigated by a local model, per the author (P3). - **Bottom line.** A short, candid practitioner recap. It supports "here is how one person works and where they draw lines" and nothing broader. ## 📖 Main content **Setup (P1-P2).** The roundtable carried the title "AI Open Space Experiences: a roundtable with speakers on their experiences with AI testing." Five people sat on stage. The author describes the format as "how we really use AI, with no bullshit." Firlej's earlier open-space exercise listed everyday uses: information gathering, functionality discovery, generating tests and code, TDD, recruitment, presentation prep, furniture design, and holiday planning. The author says the answers that follow are "roughly" what was said on stage, so they are a reconstruction. **Uses (P3).** Two work uses: analyzing codebases and test code to learn what something does, and covering teams with no dedicated quality engineer by putting the author's testing knowledge into an AI the team can query. A personal use: the author feeds decades of notes into a local open-weight model and asks what the author claims to believe and where actions contradicted it, with cited evidence. The author calls the result "a surprisingly honest mirror." The stated reason for experimenting now is access that the author expects to end. **Staying current (P4).** The author doesn't chase model releases. Sources are other testers' use of AI and coworkers, with the caveat that the author's workplace is at the front of AI adoption, so outside voices serve as a reality check. One automation: a scheduled Claude Code task that condenses about 34 AI-related Slack channels into a daily digest. **Advice and limits (P5-P7).** Advice: trust AI less, and don't assume you're far behind, because few people have tried it and foundational testing skills transfer. Limit: no AI as a proxy for talking to people, though translation and fluency help across Portuguese, Dutch, and English. Ethics: work-only use, no generated imagery that lifts from artists, and no fixed line. A Mastodon remark about humans doing menial tasks while robots make art shapes the author's wish to keep making art. **Close (P8).** The author says the value of such panels is disagreement: five daily users who differ on trust, ethics, and purpose. The post thanks the moderator and panelists. It does not say where they differed. ## 🗒️ Notes #### 📄 ¶1-¶2 Setting and format - Final day of HUSTEF 2026 in Budapest; five on stage, Firlej moderating. - Firlej ran an "AI open space" with the Polish testing community first and crowdsourced everyday uses. - The listed uses range from TDD and test generation to recruitment, furniture design, and holiday planning. - Answers are given as "roughly" what the author said. #### 📄 ¶3 What the author uses AI for - Work: codebase and test-code analysis, plus a testing-knowledge assistant for teams without a quality engineer. - Personal: a local open-weight model fed journals, therapy notes, and eight years of WhatsApp history, queried for belief-versus-action gaps with cited evidence. - Stated motive: access is unlimited now and won't stay that way, so the author wants to learn what is safe. #### 📄 ¶4 Staying up to date - Doesn't track model releases or news; follows how other testers use AI and talks to coworkers. - Coworkers are at the front of AI adoption, so outside views act as a check. - A scheduled Claude Code task digests about 34 AI Slack channels daily. #### 📄 ¶5 One piece of advice - Don't trust AI too much, and don't assume you're as far behind as you think. - Claim: almost no one has really tried it; foundational testing skills transfer. No data given. - Try it, and move on if it isn't for you. #### 📄 ¶6 What the author won't use it for - No AI as a proxy in communicating with people. - Translation and fluency help are acceptable (Portuguese, Dutch, English). - Reason: keeping the discipline of speaking to a human. #### 📄 ¶7 Ethical dilemmas - Use stays mostly within work; no imagery that lifts from artists who deserve pay. - No hard line; a sliding scale requiring repeated self-questioning. - A Mastodon line about humans doing menial tasks while robots paint and write poetry sticks with the author. #### 📄 ¶8-¶9 Closing and empty Resources heading - The author values panel disagreement on trust, ethics, and purpose; thanks Firlej, Nordström, Akehurst-Ryan, and Tuleika. - P9 is an empty "Resources" heading with no entries beneath it in the capture. ## 📚 References The link list has one entry, and it comes from the article itself. - **HUSTEF 2026**: https://hustef.com/ (anchor text "HUSTEF 2026" in the captured markup). This is the conference's own site and the primary source for the event, speaker list, and any session recording or description. **Claims with no link in the captured markup.** The roundtable title and description, the "AI open space" exercise with the Polish testing community, the Mastodon line (P7), and the panelists' identities have no link in the captured markup. The captured "Resources" heading (P9) is empty. No link points to a recording, slides, or the Mastodon post. **Primary sources worth checking for named entities.** The HUSTEF site for the session listing. Nothing else is linked, and this analysis did not fetch anything. ## 🗣️ Paraphrase At HUSTEF 2026 in Budapest, Nicole van der Hoeven sat on a five-person roundtable about how testers use AI in practice. Marta Firlej moderated and had collected everyday uses from the Polish testing community beforehand. The author reports using AI to understand codebases and test code, and to give teams without a quality engineer something that holds testing knowledge. Privately, the author runs a local model over a lifetime of notes to compare stated beliefs against behavior. The author doesn't follow release news, relies on coworkers and a daily Slack digest, and advises trusting AI less while not feeling behind. The author won't let AI write messages to people, keeps use mostly to work, avoids imagery that copies artists, and treats ethics as a sliding scale. The post says the panelists disagreed with each other but doesn't say about what. ## 📰 Newsletter **How one tester uses AI, and where they stop.** Nicole van der Hoeven's recap of a HUSTEF 2026 roundtable lists five answers. Use: reading codebases and test code, plus a testing-knowledge assistant for teams without a quality engineer. Staying current: coworkers and a scheduled Claude Code task that digests about 34 Slack channels. Advice: trust AI less, and don't assume you're behind. Limit: no AI-written messages to people. Ethics: a sliding scale, mostly work-only use. The useful part for practitioners is the explicit limits, since most AI write-ups skip them. The caveat: it's one person's account, reconstructed after the fact, and the post reports none of the other four panelists' answers. Read the original: [nicolevanderhoeven.com](https://nicolevanderhoeven.com/blog/20261008-ai-testing-roundtable-hustef/). ## ✅ Action items These items are small experiments drawn from one practitioner's account. The post gives no evidence that they work. #### DevRel - Developer advocacy: when you cite this post, say it reports one panelist's answers and not the roundtable's consensus. - Education: run an "AI open space" at a meetup with the post's format (crowdsource real uses first, then discuss) and publish the list. - Developer experience: no direct action from this source. - Developer marketing: if you quote the "no bullshit" framing (P2), attribute it to the author's description of the session. - Community management: ask your community the P4-P7 questions (what for, how to stay current, one piece of advice, never use, ethics) and compare answers. - Internal/external: no direct action from this source. #### OSPO - Policy and governance: write down your team's own "never use AI for" list, using P6 and P7 as prompts. - License compliance: no direct action from this source; nearest relevance is the P7 point about imagery that lifts from artists. - Security management: if you run a local model over personal or work notes (P3), document what data it sees and where it runs. - Software inventory: no direct action from this source. - Developer support: no direct action from this source. - Community engagement: no direct action from this source. - Strategic alignment: no direct action from this source. #### Community - Maintainers: no direct action from this source. - Contributors: no direct action from this source; nearest relevance is the P5 point that testing skills transfer to AI work. - Sponsors: no direct action from this source. - Ecosystem: no direct action from this source. #### WordPress - Core: no direct action from this source. - Core Performance: no direct action from this source. - Core AI: no direct action from this source; nearest relevance is the P3 pattern of encoding testing knowledge into an assistant a team can query. - Core Program Team: no direct action from this source. - Mobile: none. - Playground: none. - Community: none directly; see the DevRel open-space idea above. - Training: no direct action from this source; nearest relevance is the P5 claim that foundational testing skills transfer. - Marketing: none. - Support: none. - Design: none. - Documentation: none. - Photos: none. - TV: none. - Polyglots: none directly; P6 describes AI translation and fluency help as acceptable, with no evidence about quality. - Accessibility: none from this source. - Meta: none. - CLI: none. - Hosting: none. - Tide: none. - Openverse: none directly; P7 raises generated imagery that lifts from artists. - Themes: none. - Plugins: none. #### The WPOCC - Programs: no direct action from this source. - Projects: no direct action from this source. - Funding experiments: none. ### FAIR - Technical independence: no direct action from this source. - Policies and covenants: no direct action from this source. - Productized governance artifacts: no direct action from this source. ## 🧪 Hypotheses 1. A testing-knowledge assistant on a team with no quality engineer (P3) catches more missing test cases than the same team without one. 2. Testers who follow coworkers and a curated digest (P4) report as much current-practice knowledge as testers who follow release news. 3. Telling practitioners that "almost no one has really tried this" (P5) lowers their self-reported feeling of being behind. 4. A local model asked to compare stated beliefs against actions (P3) returns claims that the note-writer later rates as accurate more often than a generic personality summary. 5. Panelists who use AI daily still differ on trust and ethics (P8), and the spread is wider on ethics than on trust. ## 📊 Methodology - **H1.** Pick teams without a quality engineer. Give half a prompt-scoped assistant seeded with a testing guide. Compare defects found in a seeded-bug exercise, with the same task and time limit. - **H2.** Survey testers on their information sources and quiz them on a fixed list of current AI-testing practices. Compare scores by source type. - **H3.** Randomize a short framing statement ("few people have tried this" versus neutral) before a feeling-behind scale. Compare scores. - **H4.** Have a volunteer write a private corpus, run the belief-versus-action prompt, and rate each cited claim for accuracy and for whether the cited evidence exists. Compare with a generic summary prompt on the same corpus. - **H5.** Ask the same Likert items on trust and on ethics of a sample of daily AI users, and compare variance. This post is a single anecdote and is not the test. ## ❓ Questions - Which uses did the other four panelists name, and where did they disagree (P8)? - What did the author's testing-knowledge assistant catch or miss on teams without a quality engineer (P3)? - How does the belief-versus-action experiment keep the model from inventing evidence (P3)? - Which local open-weight model was used, and on what hardware? - What privacy steps cover therapy notes and a partner's messages (P3)? - What counts as "obviously lifting from artists" in the author's own sliding scale (P7)? - What evidence supports "almost no one has really tried this stuff" (P5)? - Is a recording of the roundtable available, and does the transcript match this reconstruction? ## 🔮 Future research - **Practitioner inventories.** Collect open-space style lists of real AI uses across testing communities and track which persist after six months. - **Self-audit prompts.** Test whether belief-versus-action prompts on personal corpora produce verifiable findings. - **Staying-current habits.** Compare digest-based and news-based habits on knowledge and anxiety. - **Ethics as a sliding scale.** Study how individual practitioners' lines move over a year. - **Access cliffs.** Document what changes for a team when unlimited AI access ends (P3). ## 💡 Implications - **Practice.** The post models stating limits next to uses. A team AI policy can copy that shape: what for, what never, and where it's unsure. - **Policy.** The P7 "no hard line" stance suggests policies need a review habit, not only a rule list. - **Security.** Feeding private notes to a model (P3) makes where the model runs a security decision. The author chose a local model. - **Education.** The P5 claim that testing skills transfer, if it holds, argues for teaching AI use as an extension of testing fundamentals. - **Funding.** The post reports no costs. The "unlimited access" remark (P3) flags that budgets may change what teams can do. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid grade of about 9 and Flesch Reading Ease of about 58, from a simple syllable heuristic on the roughly 600-word body. Question-led paragraphs and parenthetical asides lower sentence length. The estimate is rough and unverified by tooling. #### 👀 Visual learners - Present: no images or figures in the captured text. - Improve: a one-slide summary of the five questions and the author's answers would help scanning. #### 🔊 Aural learners - Present: the source was a spoken roundtable, but the capture mentions no audio or recording. - Improve: link a recording if one exists. #### 🗣️ Verbal learners - Present: each paragraph opens with the question asked (P3-P7), which makes the text easy to follow in order. - Improve: none needed at this length. #### 🧩 Kinesthetic learners - Present: no exercise. - Improve: add a prompt readers can try, such as the belief-versus-action query in P3. #### 🧠 Logical learners - Present: a clear question-and-answer structure. - Improve: the claims in P5 carry no data; a source for "almost no one has really tried this" would help. #### 🧑‍🤝‍🧑 Social learners - Present: the post names four co-panelists and credits the moderator. - Improve: report what the others said, since the post's own point (P8) is the disagreement. #### 🧘‍♀️ Solitary learners - Present: self-contained in about 600 words. - Improve: none needed. #### 🌿 Naturalistic learners - Present: none. - Improve: not applicable to this topic. ## 📊 Accessibility assessment Judged from the captured text and link list only. Rendered contrast, focus order, and alt text are unknowable from this capture. #### Screen reader semantics - Present: P9 shows a "## Resources" heading, so at least one real heading exists. The capture shows no other headings. - Improve: confirm the page has a single H1 and a logical order (WCAG 2.2 AA 1.3.1 Info and Relationships, 2.4.6 Headings and Labels). An empty Resources section is a heading with no content and may confuse navigation by heading. #### Keyboard navigation - Present: one link in the list, an ordinary anchor as far as the capture shows. - Improve: not verifiable from text. Check 2.1.1 Keyboard and 2.4.7 Focus Visible on the rendered page. #### Contrast and readability - Present: unknowable from the capture. Paragraphs are short and each opens with a bolded or plain question. - Improve: verify 1.4.3 Contrast (Minimum) on body text and link color. #### Text resizing and reflow - Present: text-only content should reflow, and no fixed-size media is in the capture. - Improve: verify 1.4.4 Resize Text and 1.4.10 Reflow at 320 CSS px and 200% zoom. #### Media and captions - No media referenced in the captured text. The text mentions a spoken roundtable but no recording, captions, or transcript. #### Forms and interactions - Present: no form in the captured text. - Improve: if a comment or subscribe form exists on the page, check 3.3.2 Labels or Instructions. #### Standards reference - Likely gaps to verify against WCAG 2.2 AA: **2.4.4** Link Purpose (In Context) for the single link, which passes if "HUSTEF 2026" names the destination in context. **1.3.1** for the empty Resources heading. Everything else needs the rendered page. ## 🌍 DEIB assessment #### Cultural breadth - Present: the event is in Budapest, the author lives across Portugal, Dutch, and English language contexts (P6), and the open-space input came from the Polish testing community (P2). - Improve: the post is in English only and reports one voice. It doesn't say what the other participants' backgrounds added. #### Inclusive language - Present: no exclusionary terms in the author's prose. P2 includes a mild profanity in a quoted framing ("no bullshit"), which is the author's own register. - Improve: none needed. #### Neurodiversity support - Present: short paragraphs, each led by the question it answers, which aids scanning. - Improve: a short list of the five questions at the top would reduce the load of tracking them in order. #### Representation - Present: five named participants. The text gives no gender or background detail and uses no pronouns for the others, so none are assigned here. - Improve: the post represents one panelist's views. Quoting the other four would widen it. #### Psychological safety - Present: the author discloses personal data uses (therapy notes, partner messages) and frames limits without judging others. The post says the panelists land "in very different places" and treats that as the value (P8). - Improve: the partner's messages (P3) are described without noting consent. The post doesn't say whether the partner agreed. ## ✅ Logical validity #### Strong reasoning - **Stated limits, ¶6-¶7.** The author names what AI won't be used for and admits the ethics line is a sliding scale, which is internally consistent with the "keep questioning myself" stance. - **Source caveat, ¶4.** The author notes that the workplace is at the front of AI adoption and that outside voices are a reality check, which flags a bias in the author's own sample. - **Privilege disclosure, ¶3.** The author says access is effectively unlimited now and won't last, which tells readers the practice may not generalize. - **Reconstruction disclosure, ¶2.** "Roughly how I answered" marks the answers as recalled, not transcribed. #### Fallacies No formal fallacies found. The post reports personal practice and advice and doesn't argue a thesis. What I checked: whether claims exceed their evidence and whether the headline matches the body. Points on evidence: - **Unsupported generalization, ¶5.** "Almost no one has really tried this stuff." No survey or number appears, and the claim covers the whole field. *Fix:* attribute it to the author's observation ("most people I talk to"). - **Single-source framing, title.** "How We Really Use AI in Testing" promises the group's practice, while the body reports only the author's answers. *Fix:* "How I really use AI in testing, from a HUSTEF roundtable." - **Open question, ¶3.** The belief-versus-action result is called "a surprisingly honest mirror" with no check that the cited evidence was accurate. *Fix:* say how the cited evidence was verified. No instructions addressed to an AI or reader appear in the source text or link list. #### Tone of voice - Casual, candid, and first-person, in a conference-recap register with contractions and a parenthetical aside. This is opinion and practitioner experience. The author has no stated commercial interest in the captured text, though the author names Claude Code as a tool in use (P4), so readers should note the author's tool preference. Viewpoint: the author is pro-use with explicit limits and doesn't argue for adoption. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response to accusation, and this post responds to no accusation. Each prong is still checked. - **Deny.** None. The author answers moderator questions and denies nothing. - **Attack.** None. No person or group is criticized. The author describes disagreement among panelists (P8) without naming a target. - **Reverse victim and offender.** None. The author describes personal privilege (P3) and ethical self-questioning (P7), with no claim of being wronged. ## 🎯 Conclusion The post is a short first-person recap of one panelist's answers at a HUSTEF 2026 roundtable. It reports specific habits: AI for reading code and standing in for a missing quality engineer, a local model over private notes, coworkers and a daily digest instead of release news, and firm limits on AI-written messages and on imagery that lifts from artists. What it doesn't establish: whether any of these practices work, what the other four panelists said, or how the claim that "almost no one has really tried" AI holds up. The title promises "we" and delivers "I." Use it as one practitioner's account. **Next steps for readers:** try the open-space format with your own community, write down your own never-use list, and check the HUSTEF site for a recording of the session. This analysis describes the post and doesn't judge the author's choices. ## 🔄 Refactor Appendix — Roundtable: How We Really Use AI in Testing (HUSTEF 2026) > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 1 Permanent. Resonance filter applied: the panel logistics > and thank-yous were dropped as event detail. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Trust AI less and don't feel behind > [!quote] [¶5] > Don’t trust AI too much, and don’t believe you’re as far behind as you think. **Context & use**: - ¶5, the author's one piece of advice at the roundtable. - The "far behind" half rests on an unsupported claim that almost no one has tried AI. Treat it as opinion. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H2] Never let AI write your messages to people > [!quote] [¶6] > I never want it to craft a whole message I didn’t think of. **Context & use**: - ¶6, the author's single hard limit. Translation and fluency help stay allowed. - The stated reason is keeping the discipline of talking to a human. **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> #### [H3] Encode testing knowledge into a team assistant > [!quote] [¶3] > baking my testing knowledge into an AI the team can ask questions of **Context & use**: - ¶3, a work use for teams with no dedicated quality engineer. - The post reports no results, so the benefit is a claim of practice, not evidence. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H4] A scheduled digest replaces chasing AI news > [!quote] [¶4] > a scheduled Claude Code task that scrapes my ~34 AI-related Slack channels into a daily digest **Context & use**: - ¶4, the author's alternative to following model releases. - A concrete automation pattern: many channels in, one filtered daily output. **Tags**: #xref #addendum <!-- atomicity: 1 idea ✓ --> #### [H5] AI ethics is a sliding scale > [!quote] [¶7] > There’s no hard line; it’s a sliding scale, and I have to keep questioning myself. **Context & use**: - ¶7, the author's answer on ethical dilemmas, tied to avoiding imagery that lifts from artists. - Pairs with the Mastodon line about robots making art, which the author reports but doesn't link. **Tags**: #xref #bmf <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] State AI limits beside AI uses **Claim** (one sentence, Courtney's voice): A useful AI practice write-up names what the person won't use AI for, so readers can see where the edges are and not only the wins. **Evidence pointer**: [[#H2]] · [[#H5]] · [[#H1]] **Scope & conditions**: Holds for personal and team practice accounts. Weakens when the limits are vague or the account reports one person's view as the group's. The source reports no outcomes, so this is a writing-practice claim, not an effectiveness claim. **Links**: - [[Team AI policy]] — because applies-in a never-use list - [[Practitioner AI inventories]] — because extends - [[Headline attribution]] — because contrasts on a title that says "we" over one voice <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Look for a HUSTEF 2026 recording of the AI roundtable and compare it with this recap — https://hustef.com/ #fleeting #inbox - [F2] Draft my own never-use-AI list, using the P6-P7 questions as prompts #fleeting #inbox - [F3] Open question: how would a belief-versus-action prompt on my own notes avoid invented evidence? #fleeting #inbox - [F4] Article seed: run an "AI open space" at a meetup before the talks #fleeting #inbox #articleseed - [F5] Note the daily-digest pattern from P4 for my own Slack channels #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Check hustef.com for a recording or session notes for the AI roundtable — research, 10 min #things - [ ] Write my own five-line "what AI is for, what it's never for" note — personal practice #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Roundtable How We Really Use AI in Testing.md|Roundtable: How We Really Use AI in Testing by Nicole van der Hoeven]] ## Highlights - What do you use AI for? Mostly two things at work: analysing codebases and test code to understand what something actually does, and filling gaps on teams without a dedicated quality engineer — baking my testing knowledge into an AI the team can ask questions of. On the personal side, I’ve run some weird experiments: I’m a lifelong note-taker with decades of context (journals, therapy notes, eight years of WhatsApp history with my partner), so I’ve fed it all into a local, open-weight model and asked things like “what do I claim to believe, and where have I acted contrary to that — cite your evidence?” It’s a surprisingly honest mirror. I’m in a privileged spot with effectively unlimited AI access right now, and I know it won’t last, so I want to find out what’s safe and what isn’t while I can. ([View Highlight](https://read.readwise.io/read/01m4rafccjztta09427xa73snx)) - Note: [H3] Encode testing knowledge into a team assistant - Tags: atom/literature/h3 - How do you stay up to date? Honestly, I don’t — not by chasing model releases or news. I keep up with how other people, especially in testing, are using AI, and I lean on my coworkers (we’re at the forefront of this at work, so stepping outside that bubble is a useful reality check). I do have a scheduled Claude Code task that scrapes my ~34 AI-related Slack channels into a daily digest of just the things I care about. ([View Highlight](https://read.readwise.io/read/01m4rafejkqm64gtat3j8a3dzt)) - Note: [H4] A scheduled digest replaces chasing AI news - Tags: atom/literature/h3 - One piece of advice? Don’t trust AI too much, and don’t believe you’re as far behind as you think. Almost no one has really tried this stuff; it’s genuinely early, everyone’s learning, and your foundational testing skills transfer. Just pick it up, play, and if it’s not for you, move on. ([View Highlight](https://read.readwise.io/read/01m4raf7b4v41q95xhwmka99js)) - Note: [H1] Trust AI less and don't feel behind - Tags: atom/literature/h1 - Anything you’d never use AI for? Communicating with people as a proxy. I’ll use it to translate or to help me sound more fluent (I live across multiple languages — Portugal, Dutch, English), but I never want it to craft a whole message I didn’t think of. I don’t want to lose the discipline of actually talking to a human. ([View Highlight](https://read.readwise.io/read/01m4raf9y4b7081mnc4ve6apj2)) - Note: [H2] Never let AI write your messages to people - Tags: atom/literature/h2 - Ethical dilemmas? Yes — which is exactly why I keep my use mostly in the context of work, and try not to generate imagery or anything that’s obviously lifting from artists who deserve to be paid for it. There’s no hard line; it’s a sliding scale, and I have to keep questioning myself. (A Mastodon line stuck with me: the strange future where the humans do the menial tasks and the robots write poetry and paint. I try not to let AI stop me from making art.) ([View Highlight](https://read.readwise.io/read/01m4rafgwgmtfd1w9re6y02r4n)) - Note: [H5] AI ethics is a sliding scale - Tags: atom/literature/h3