How AI is changing what counts as evidence of human capability
Signal Intelligence Briefing 007
Edge line: When competent output becomes easy to produce, institutions have to decide what still proves capability.
For a long time, many systems relied on a simple inference.
A person produced an artefact. The artefact looked competent. The institution inferred something about the person.
An essay suggested knowledge. A portfolio suggested skill. A coding task suggested programming capability. A report suggested professional judgement. A polished answer in an interview suggested understanding.
That inference was never perfect. It has always depended on context, identity, authorship, task design and trust. But generative AI has weakened it enough that institutions are beginning to redesign the evidence around it.
In Aotearoa New Zealand, this is no longer only a research question. Waipapa Taumata Rau / University of Auckland has formally adopted a two-lane assessment approach: controlled assessment where AI is restricted by default, and uncontrolled assessment where AI use is permitted while the student remains responsible for the work. NZQA advises schools and kura to use milestones, observation and follow-up conversations to establish authenticity. Massey, Auckland and Victoria have all stepped back from relying on AI-detection software as decisive evidence. The issue has also crossed into mainstream public discussion. In August 2026, Stewart Sowman-Lund’s Sunday Star-Times / The Post reporting compared the different ways New Zealand universities were responding to AI and described Auckland’s two-lane model, including the possibility that students may need additional measures to demonstrate understanding.
The structural question underneath these responses is larger than cheating.
If a finished output no longer tells us enough about the person behind it, what should count as credible evidence of human capability instead?
1. Purpose of This Briefing
Primary Goal
To examine the emerging shift in how institutions establish credible evidence of human capability when generative AI can materially contribute to competent-looking artefacts and performances.
Core Strategic Premise
The central problem is not whether AI was used. It is what the resulting evidence can legitimately establish about the human presenting it.
As AI becomes a normal part of some real-world tasks, systems increasingly need to distinguish at least four questions: Is this the person they claim to be? Did they produce or direct this work? What can they do independently? What can they responsibly accomplish with AI?
Those are related questions. They are not the same question.
What This Briefing Is Not
This is not an article about AI cheating. It is not a case for handwritten examinations, surveillance-heavy assessment or oral examination everywhere. It does not claim that finished artefacts have become meaningless. Nor does it claim that every employer or professional body is already redesigning capability verification.
The evidence is strongest in education and assessment. Outside education, the pattern is emerging and concentrated in selected technical and analytical roles where AI can perform much of the very task that conventional assessment was intended to test.
Desired Reader Outcome
Readers should leave able to distinguish verification architecture from capability architecture, recognise the trade-offs in each, and ask a more precise question of their own systems:
What does this evidence actually prove?
2. Executive Signal Summary
Signal 1 — Some conventional outputs are becoming weaker standalone evidence of independent human capability
Generative AI can materially improve the apparent competence of written, analytical and technical outputs. That does not make the output valueless. It does mean that the relationship between the quality of the artefact and the independent capability of the person presenting it is less direct than many systems previously assumed.
Signal 2 — Education systems are already redesigning assessment architecture
Auckland, Sydney and LSE are using combinations of controlled, secure or observed assessment alongside assessment in which AI use is permitted. NZQA recommends milestones, observation and follow-up conversation. TEQSA is explicitly treating generative AI as a learning-assurance and assessment-design problem rather than simply a detection problem.
Signal 3 — AI detection has not reliably restored evidential trust
Some institutions have reduced or abandoned the use of AI-detection software because the evidential standard is too weak for consequential decisions about misconduct. This does not remove the need for authenticity. It pushes the problem back towards assessment design, professional judgement and additional evidence.
Signal 4 — A narrower version of the same problem is appearing in employment
Google has publicly linked a return to in-person technical interviewing with the need to verify fundamentals when candidates can use AI during remote coding tasks. GitLab is piloting a redesigned technical interview process that explicitly allows AI while shifting attention towards design judgement, code comprehension, validation and reasoning. McKinsey has reportedly piloted AI-enabled final-round business cases in which candidates are judged on how they use and interpret AI-supported outputs.
Signal 5 — The emerging choice is not simply AI or no AI
The more mature design problem is whether to establish baseline independent capability, AI-enabled capability, or both. Two-lane models are one emerging response: protect some independent demonstration while also assessing how people direct, evaluate, adapt and take responsibility for AI-assisted work.
3. Why This Matters
Trust depends on what evidence can bear
Credentials, hiring decisions, professional registration and assessment all depend upon inference. A certificate does not contain competence. A portfolio does not contain judgement. An exam script does not contain understanding. Institutions infer those things from evidence gathered under particular conditions.
When the conditions change, the inference may need to change with them.
Generative AI matters because it reduces the cost of producing plausible competence. Someone with limited knowledge can now produce a strong-looking paragraph. A novice programmer can produce functioning code. A candidate can rehearse or generate polished answers. A competent professional can also use AI to work more quickly and at a higher level.
These outcomes are not equivalent. The artefact alone may no longer tell us which one occurred.
That creates a trust problem even where nobody has cheated.
Capability is not the same as authorship
One institutional temptation is to collapse the issue into provenance: who wrote this?
That matters. But authorship is only one layer.
A student may legitimately use AI and still understand the work deeply. Another may technically write every word themselves but reproduce a memorised method without being able to adapt it. A professional may direct AI to produce an excellent report while exercising substantial judgement over what is accepted, rejected and changed. Another may submit fluent output they cannot meaningfully defend.
Knowing who produced the text does not automatically tell us what the person can do.
Equity can be damaged by bad verification
The obvious response to uncertainty is more control: invigilate more, watch more closely, check identity more aggressively, demand oral defence, record process, track keystrokes.
Some of that may be justified in high-consequence settings. But verification is not free of bias or burden. Oral assessment can disadvantage people for reasons unrelated to the capability being tested. Surveillance can create accessibility, privacy and trust problems. Controlled conditions can measure a narrower construct than the real-world task. Process tracking can turn legitimate learning variation into suspicion.
The challenge is not to verify everything. It is to use the minimum evidence necessary to support the decision at stake.
Workforce capability is changing shape
In roles where AI becomes part of legitimate work, competence may increasingly include the ability to direct, evaluate, challenge and take responsibility for AI-supported outputs.
This is not merely “AI literacy”. It is closer to supervisory judgement: knowing what to ask for, recognising when the output is weak or dangerous, adapting it to context, checking what matters and remaining accountable for the result.
That may become especially important as AI absorbs more of the visible production layer. The human contribution does not disappear. It moves towards framing, evaluation, contextual judgement and responsibility.
4. Current System Reality
Auckland has moved beyond a simple prohibition model
Waipapa Taumata Rau / University of Auckland now provides one of the clearest New Zealand examples of the emerging architecture.
Its Two-Lane Approach distinguishes controlled assessment, where AI is restricted by default, from uncontrolled assessment, where AI use is not restricted. The university explicitly connects this model both to assurance that students have achieved learning outcomes and to the development of critical, discerning and responsible AI capability. Programme-wide implementation is required by 2027.
This is more significant than a rule about whether ChatGPT is allowed.
It acknowledges that a single assessment condition may no longer do all the work. Some capabilities need to be demonstrated under controlled conditions. Others should be developed and assessed in the conditions in which contemporary practice actually occurs.
The model is not perfectly binary. Auckland allows assessment designs to combine controlled and uncontrolled elements, and encourages programme-level thinking about AI capability across the whole learner journey.
That matters. The strongest emerging signal is not two boxes. It is the recognition that institutions may need more than one evidential lane.
New Zealand public debate is catching up with institutional change
Stewart Sowman-Lund’s August 2026 Sunday Star-Times / The Post article made the divergence between university responses publicly visible. It compared Auckland’s two-lane model with different approaches at Massey and Victoria and quoted university staff describing how AI has disrupted confidence in submitted work as a proxy for learning.
The article also surfaced a practical boundary that matters for this briefing. Additional evidence may include oral presentations, in-class activity or direct engagement, but scale remains a constraint. A ten- or twenty-minute professional conversation may be entirely sensible for a high-stakes decision. It is a different proposition for every student in a class of two hundred.
The public discussion has therefore moved beyond whether universities “allow AI”. The emerging question is what kinds of evidence remain credible, proportionate and educationally useful.
NZQA is pointing towards conversation and observation
NZQA’s guidance to schools and kura recommends authenticity strategies that include milestones, observation, source identification and follow-up conversations. It explicitly suggests asking learners how they reached a conclusion.
That is a small but important change of emphasis.
The final artefact remains part of the evidence. But the learner’s ability to explain how they arrived there becomes evidence too.
I use the term professional conversation deliberately. A viva is one academic form of oral verification. Professional conversation is broader. It can be verification-oriented, but it can also reveal judgement, reflection, adaptation and transfer.
In Aotearoa, this also opens a design question rather than an answer. Professional conversation need not be assumed to mean only the conventional academic viva. In appropriate contexts, future work might explore whether forms of kōrero or talanoa can contribute to capability evidence — but only with the relevant cultural grounding and authority. They should not be treated as interchangeable labels for oral assessment. For now, the point is simply to leave that design space open.
Detection is proving insufficient as an evidential shortcut
One of the strongest signals is a failed solution.
Massey stopped using AI-detection software. Auckland had already decided not to endorse or use it as an evidential basis. Victoria did not use it. The common concern was reliability.
This matters because detection was the obvious technical answer to a technical problem: if AI helped produce the artefact, use another system to detect the AI.
But a detector can only ever answer a narrow question, and often imperfectly. It cannot establish understanding. It cannot reliably distinguish borrowed fluency from genuine judgement. It cannot show whether a learner can reproduce the capability in another context.
When detection is too weak for consequential decisions, the system is forced back towards stronger evidence design.
Sydney and LSE show comparable international movement
The University of Sydney also uses a secure/open assessment model. Secure assessment establishes knowledge, skills and attributes without external aids. Open assessment permits contemporary tools, including AI, and supports responsible use.
LSE introduced an Observed Assessment policy in 2025. It requires programmes to include appropriate observed safeguards and recognises a range of forms: in-person examinations, oral assessment, in-class activity, linked formative and summative work, and some technology-enhanced process evidence.
These institutions are not converging on one method. They are converging on a problem definition: unsupervised output alone cannot always assure the learning outcome it is supposed to represent.
TEQSA is making the system-level case
Australia’s tertiary regulator, TEQSA, has framed generative AI as a learning-assurance problem requiring assessment reform. Its guidance emphasises learning outcomes, evaluative judgement, critical thinking and ethical reasoning, and documents approaches including secure tasks, process evidence and writing history.
The significance is not that Australia has discovered the perfect assessment model. It is that the regulatory conversation has moved from “detect the AI” towards “design evidence that can support the assurance claim”.
The same pattern is beginning to appear outside education — but narrowly
The employment evidence needs more restraint.
This is not yet a general labour-market shift. Much employer activity is about identity fraud, proxy candidates and deepfakes rather than capability design. Those are real problems, but they should not be used to inflate the argument.
A smaller cluster is more interesting.
Google’s chief executive has publicly linked the return of some in-person technical interviews with the need to ensure candidates still possess computer-science fundamentals in an environment where AI can assist remote coding tasks.
GitLab’s AI-Native Hiring Working Group is more explicit. Its proposed interview redesign acknowledges that AI coding assistants can perform much of what traditional technical interview tasks measure. The proposed response is not to ban AI from every stage. It is to observe design judgement, communication of intent, code comprehension, iterative problem solving and validation of AI-supported outputs.
McKinsey has reportedly piloted a similar logic in selected final-round graduate interviews by requiring candidates to use its internal AI tool and evaluating how they interpret and apply the result.
The footprint is small. But the structural rhyme with education is difficult to ignore.
5. What the System May Be Misunderstanding
Misunderstanding 1 — The problem is AI use
AI use may be permitted, prohibited or required depending on the task.
The deeper question is validity: what capability is being claimed, and what evidence would justify believing that claim?
A rule that says “AI was used” tells us very little by itself. The same use may be misconduct in one assessment, responsible professional practice in another and the explicit learning objective in a third.
Misunderstanding 2 — Detecting AI restores authenticity
Detection can be a risk signal. It is not the same as evidence of understanding or competence.
Even perfect authorship detection would leave another question unanswered: what can the person do?
A system can authenticate an artefact and still measure the wrong thing.
Misunderstanding 3 — Secure assessment automatically measures capability better
Secure conditions can establish something important: unaided or bounded performance under known conditions.
But real-world capability may involve tools, collaboration, reference material and AI. A closed task can therefore strengthen assurance about fundamentals while simultaneously narrowing ecological validity.
The stronger question is not secure or open. It is which capability belongs in which condition.
Misunderstanding 4 — AI-enabled work is less human
Human contribution should not be measured only by the number of words typed or lines of code written.
In some contexts, the more consequential human work lies in defining the task, recognising error, comparing alternatives, adapting to context, exercising professional judgement and accepting responsibility.
If AI becomes part of legitimate practice, capability architecture must learn to see those forms of contribution.
Misunderstanding 5 — One assessment lane can do everything
The old model often asked one artefact to carry several inferences at once: authorship, knowledge, process, judgement and capability.
That may now be too much weight for one object.
A two-lane or multi-evidence architecture does not need to become bureaucratic. It simply recognises that different claims may require different evidence.
6. Emerging Adaptation Patterns
Pattern A — Restore controlled conditions
In-person, supervised, observed or otherwise controlled tasks are being used to establish baseline independent capability, identity or authorship.
This is the most conservative response. It preserves an older evidential relationship by restoring the conditions under which the artefact was originally trusted.
Pattern B — Add professional conversation
Follow-up questioning, oral assessment, viva-style defence and professional conversation test whether the person can explain, interrogate and take responsibility for the work presented.
The distinction matters. When used for verification, a viva may ask, in effect, “Can you demonstrate that you understand and can account for this work?” A professional conversation can ask a richer set of questions: What did you notice? What did you trust? What did you reject? What would change in a different context? What are you responsible for here?
In some fields, that conversation may become one of the clearest ways to make judgement visible.
Pattern C — Capture process and provenance
Milestones, draft history, working records, source trails and process evidence supplement the final artefact.
This approach recognises that capability may be distributed across a sequence rather than contained in the end product. It can reveal development, correction and decision-making. It can also become intrusive if institutions begin treating every click as evidence.
Pattern D — Permit AI and raise the construct
Some organisations are allowing AI while changing what they assess.
Instead of rewarding raw production, they look for critique, validation, debugging, adaptation, explanation and responsible use. The person is not asked to compete with the machine at machine production. They are asked to demonstrate command over the human–AI work system.
Pattern E — Split the architecture
The most interesting emerging response combines independent and AI-enabled demonstration.
One lane asks: what can this person do without unverified assistance?
The other asks: what can this person responsibly accomplish with contemporary tools?
This is not appropriate everywhere. But it is a more coherent response than pretending that AI can either be excluded from all meaningful work or accepted without changing the evidence model.
Pattern F — Describe capability as a progression, not a binary
A further implication is beginning to emerge from field practice.
AI capability is unlikely to be well described by a binary such as “can use AI / cannot use AI”, or by prompt skill alone. One provisional six-step progression I am currently exploring in tertiary capability work begins with access and orientation, then moves through purposeful use, critical evaluation, contextual and relational adaptation, responsible practice, and finally accountable judgement and transfer.
This is not a validated national framework. It is a working model. Its value for this briefing is conceptual: once institutions stop treating output quality as sufficient evidence, they need language for how increasingly sophisticated AI-enabled human capability might actually develop.
7. Strategic Implications
For educators and assessment designers
The first design question should be: which capability must the learner be able to demonstrate independently, and which capability should include contemporary tools?
That decision belongs before the assessment format.
If independent recall, calculation, physical performance or foundational reasoning matters, establish it directly. If professional practice genuinely includes AI, then create evidence of how the learner directs, evaluates and takes responsibility for AI-supported work.
An authenticity policy is not a substitute for assessment design.
For employers
Reconsider what familiar hiring artefacts now reveal.
A take-home task may still be useful. A coding test may still be useful. A portfolio may still be useful. But their evidential value should not be assumed merely because the format is familiar.
Where AI can perform much of the task, the employer may need either a controlled baseline check, a live professional conversation, an AI-enabled practical demonstration, or some combination.
The goal is not to catch candidates out. It is to understand what capability the organisation is actually buying.
For credentials and recognition systems
Recognition systems have always faced a version of this problem. Qualifications, RPL, workplace evidence and professional registration all depend on connecting claims of competence with defensible evidence.
AI increases the pressure on that connection.
A credential becomes more trustworthy when there is a clear relationship between what it claims and what the holder has demonstrably done, explained, adapted or reproduced. That may strengthen the case for richer workplace evidence and professional conversation — especially for capabilities that are already difficult to capture through conventional written assessment.
For professional bodies
Professional competence increasingly includes responsibility for tool-supported work.
Some professional guidance is beginning to emphasise that practitioners must understand enough about AI-supported outputs to verify them and remain accountable for their use.
That is not yet the same as redesigned capability assessment. But it points towards a future in which professional authority cannot be separated from the ability to supervise synthetic contribution responsibly.
For leaders and policy makers
The institutional choice is not simply permissive versus restrictive AI policy.
Leaders need an evidence architecture: a defensible account of which claims require strong identity or authorship assurance, which require independent capability, which may legitimately be AI-enabled, and what burden of verification is proportionate to the consequence.
The wrong response could be expensive and invasive. The absence of a response could leave credentials and hiring decisions resting on evidence that no longer means what the institution thinks it means.
For culturally intelligent capability design
In Aotearoa, professional conversation should not default automatically to an imported viva model.
The more careful next question is whether future capability-evidence methods can be designed with Māori and Pacific relational practices in view, where appropriate and with the relevant authority. Kōrero and talanoa are possibilities for exploration, not labels to retrofit onto an existing assessment method.
8. Intervention Options
Option 1 — Harden the existing artefact
Use secure conditions, identity checks, detection tools or supervision to preserve the existing assessment object.
Suitable when independent performance is central and the consequence is high.
Trade-off: simple to govern, but may increase surveillance and may measure a narrower capability than authentic practice requires.
Option 2 — Add evidence around the artefact
Retain the existing output but supplement it with professional conversation, observation, process records, source trails or live demonstration.
Suitable when the artefact remains useful but no longer carries enough evidential weight by itself.
Trade-off: richer evidence, but more time, cost and potential accessibility or bias concerns.
Option 3 — Redesign the capability construct
Explicitly permit AI where appropriate and assess judgement, adaptation, validation, explanation and accountable use.
Suitable when AI is part of legitimate contemporary practice.
Trade-off: more authentic, but harder to standardise and dependent on clearer definitions of what human capability now means.
Option 4 — Use a two-lane architecture
Require a bounded independent demonstration of foundational capability alongside AI-enabled assessment of higher-order practice.
Suitable when both independent fundamentals and tool-enabled professional performance matter.
Trade-off: potentially stronger evidence, but operationally more complex and unnecessary in low-risk settings.
Option 5 — Do nothing yet
Where the artefact remains a valid proxy and the consequence is low, avoid redesign for its own sake.
Suitable when AI does not materially affect the capability being inferred.
Trade-off: preserves simplicity, but only if the institution has deliberately tested the assumption that its existing evidence still means what it thinks it means.
9. Emerging Directional Principles
1. Assess the construct, not the anxiety.
2. Distinguish identity, authorship, independent capability and AI-enabled capability.
3. Use the minimum verification necessary for the risk and consequence involved.
4. Do not treat surveillance as a synonym for validity.
5. Preserve human judgement where automated evidence is weak or contested.
6. Where AI is legitimate to the real task, assess the quality of human direction, evaluation, adaptation and accountability.
7. Where independent fundamentals matter, require them to be demonstrated rather than inferred from polished output.
8. Prefer complementary evidence points when one artefact carries too much inferential weight.
9. Treat professional conversation as a capability method, not merely an anti-cheating device.
10. In Aotearoa, leave room to explore culturally grounded relational evidence rather than assuming one imported oral-assessment form is universally valid.
10. Watchlist
The interpretation remains conditional. The following developments would strengthen, weaken or materially redirect it.
• Whether two-lane assessment models spread beyond tertiary education and technical hiring.
• Whether New Zealand universities’ current experiments converge or continue to diverge materially.
• Whether more employers explicitly allow AI during assessment while raising the capability construct.
• Whether professional licensing bodies redesign assessment instruments rather than only issuing AI-use guidance.
• Whether reliable provenance technologies emerge that reduce the need for live verification.
• Whether professional-conversation, oral and observed assessment expansion produces measurable validity, accessibility or equity problems.
• Whether process-tracking tools become accepted evidence or create new surveillance burdens.
• Whether New Zealand employers or professional bodies begin separating independent and AI-enabled capability explicitly.
• Whether AI capability frameworks begin describing progression through evaluation, contextual adaptation, responsibility, judgement and transfer rather than tool access or prompt fluency.
• Whether institutions eventually retreat from secure assessment because AI-enabled practice becomes too central to exclude.
• Whether culturally grounded forms of conversational evidence are developed with sufficient authority and validation to become credible alternatives or complements to conventional oral assessment.
11. Conclusion
Generative AI has not made human capability unknowable.
It has exposed how much institutional confidence rested on proxies whose meaning was never as stable as it appeared.
A good essay used to suggest that the student could think. A polished application suggested that the candidate could communicate. A successful coding task suggested that the programmer could code. Those inferences still sometimes hold. But where AI can materially produce the visible performance, the institution needs to know more about the conditions under which the evidence was produced and the capability it is supposed to represent.
The weak response is to trust the polished artefact anyway.
The equally weak response is to answer every uncertainty with surveillance.
A stronger response begins by separating the claims.
Who is this person? What did they author or direct? What can they do independently? What can they accomplish responsibly with AI? What evidence would be proportionate to the decision at stake?
That may lead to secure assessment in some places. Professional conversation in others. Process evidence, workplace demonstration or AI-enabled tasks elsewhere. In high-consequence settings it may require more than one lane.
The deeper shift is not from written assessment to oral assessment, or from open to closed conditions.
It is from assuming that the artefact speaks for the capability to deliberately designing evidence that does.
12. References
Waipapa Taumata Rau / University of Auckland. Advice for students on using Artificial Intelligence.
Waipapa Taumata Rau / University of Auckland. Assessment of Courses Procedures.
Waipapa Taumata Rau / University of Auckland TeachWell. Two-Lane Approach to assessment.
NZQA. Guidance on the acceptable use of Artificial Intelligence.
University of Sydney. Assessments.
London School of Economics and Political Science. School position on generative AI.
TEQSA. Enacting assessment reform in a time of artificial intelligence.
TEQSA. Principles for criteria and standards in assessment for gen AI use.
GitLab Handbook. AI-Native Hiring Working Group.
The Guardian. “McKinsey asks graduates to use AI chatbot in recruitment process.” 14 January 2026.
Social Workers Registration Board. Overseas-qualified social workers / GenAI screening guidance.
Engineering New Zealand. Assessment guidance for CPEng.
13. Key Questions
When does a finished artefact cease to be sufficient evidence of capability?
When the relationship between the quality of the output and the capability being inferred becomes too weak or ambiguous for the consequence of the decision. AI does not automatically invalidate an artefact. It changes the confidence institutions can place in it under some conditions.
What is the difference between proving authorship and proving capability?
Authorship asks who produced or directed the work. Capability asks what the person can understand, judge, explain, perform, adapt or reproduce. Authenticating authorship may strengthen evidence, but it does not by itself establish capability.
Which capabilities should still be demonstrated without AI?
Those for which independent performance is itself consequential: foundational knowledge or skills, safety-critical actions, professional fundamentals, or capabilities people must retain when tools are unavailable or wrong. The answer is domain-specific rather than ideological.
What should count as credible evidence of AI-enabled human capability?
Evidence that reveals how the person directs the task, evaluates output, checks important claims, adapts to context, explains decisions, recognises limitations and remains accountable for the result. A polished AI-supported artefact alone is usually weak evidence of those capabilities.
Is professional conversation the answer?
Sometimes. It can reveal understanding, judgement and transfer that a static artefact cannot. It is also costly, can be biased and may be inappropriate at scale. It should be one method within a wider evidence architecture rather than a universal replacement for written assessment.
When does verification become surveillance?
When evidence collection exceeds what is proportionate to the capability, risk and consequence being assessed; when monitoring becomes continuous rather than purposive; or when the system collects behavioural traces without a clear validity case.
Is a two-lane model necessary everywhere?
No. It is most useful where both independent fundamentals and AI-enabled professional capability matter. Low-risk tasks or domains in which AI does not materially alter the evidential relationship may not require it.
14. Continue Exploring
SIGNAL 004 — When Answers Become Abundant, Learning Evidence Must Change
SIGNAL 005 — When Domain Expertise Becomes Executable
This issue extends the earlier question of evaluative judgement and technical amplification into the evidence problem: if AI contributes materially to competent work, what human capability remains important to see and verify?
SIGNAL 006 — The Human Operating Envelope
This issue also connects to the distinction between assisted task performance and durable human capability. Output can improve while what the person can later explain, adapt or reproduce remains uncertain.
Visible Thinking and capability evidence remain part of this wider landscape, but SIGNAL 007 is not an argument for any single tool. The deeper question is what evidence architecture becomes necessary when output quality can no longer carry the inference alone.
15. Subscribe
Signal Intelligence examines structural changes in AI, capability, education, assessment, trust and institutional design.
Subscribe to Signal Intelligence
16. Tags
AI; capability; assessment; evidence; trust; credentials; workforce; human judgement; professional conversation; AI capability; learning assurance; Signal Intelligence
Kia ora! Hey, I'd love to know what you think.