Choose one recurring task and identify the judgment it requires. Use independently authored or public material only when its reuse terms permit the intended project and audience; Creative Commons explains that conditions vary by license. Build one bounded AI step, show the cases and human checks, and limit your claim to what they demonstrate. Redaction or synthetic data do not by themselves establish permission, privacy, or fidelity: NIST SP 800-226 describes possible utility challenges for synthetic data, not inevitable outcomes. If permitted inputs cannot preserve meaningful task judgment, change tasks or seek specific approval for a bounded private demo.
What is the smallest project that can answer the question?
Choose one recurring task whose useful judgment can be demonstrated with material you are permitted to use in the tools and for the audience you have in mind. Then build the smallest inspectable test: identify the decision the task supports, isolate one AI operation that might assist it, and show enough of the input, output, and human review for another person to understand what happened. If a safe input removes the judgment that makes the task meaningful, choose a different task or seek specific approval before using the relevant context. This is a practical decision rule synthesized from the assignment’s constraints, not a validated hiring recipe.
Start by writing the task as an action, not a job title: for example, “sort incoming requests into a small set of next steps,” rather than “automate my analyst role.” State what the human currently decides, what evidence informs that decision, and where an AI system would enter. Keep the first version to one operation, such as proposing a category or extracting a field. That boundary makes it possible to inspect whether the operation helps with the stated decision without suggesting that an entire occupation can be represented by a short demo.
Next decide what a viewer could actually examine. A useful demonstration might expose a permitted sample, the AI’s suggestion, the rule or evidence used to check it, and the human’s final decision. The viewer should be able to distinguish the model’s contribution from the parts you designed, checked, or corrected. If you cannot show the input, describe its source and limits at a level you are allowed to disclose, then avoid claiming that unseen details establish the result. The project’s purpose is to make one piece of reasoning legible; visual similarity to a former workplace process is not itself evidence that the demonstration captures that process.
This also sets a boundary around the conclusion. A portfolio project can show how a bounded workflow was constructed and what happened on the material it presents. It does not, by itself, establish that the workflow is suitable for deployment in an organization, that it works reliably on other cases, or that its maker is ready for a particular job. Those conclusions require evidence beyond a small demonstration. A close replica may seem more role-specific, but reproducing protected context or implying unshown representativeness would make the project harder to explain honestly.
Can I use public material just because I can access it?
No. A public page or downloadable file is not automatically available for every reuse, modification, portfolio audience, or AI-tool workflow. Before choosing it, record the item’s title or identifying description, its creator or rights holder if stated, its source URL, the license or terms attached to that item, any changes you plan to make, the intended viewers, and whether you will submit it to an AI service. This source register gives you a traceable reason for choosing the material; it is not a legal determination that all rights or uses are cleared.
The Creative Commons page “Creative Commons Licenses” explains why the license matters: the six licenses combine conditions such as attribution, limits on commercial use, restrictions on adaptations, and share-alike requirements in different ways. So a search result, an open webpage, or a “free to download” label does not answer the relevant question. Follow the source to the item-level notice or terms, and record the exact license and version where one is supplied. If a page contains several items, check whether the notice applies to the particular text, image, dataset, or other component you want to use rather than assuming one label covers everything on the page.
Then compare the conditions you recorded with the project you actually plan to make. Is the portfolio public or limited to a named reviewer? Will you adapt, translate, combine, or redistribute the material? Will you use it commercially, including as part of a portfolio connected to paid work? Will an AI provider process the input, and do that provider’s and source’s terms fit the intended use? These are practical questions for defining the project’s use, not a substitute for reviewing applicable terms or getting qualified advice where the consequences matter. If the terms do not clearly accommodate the planned use, do not infer permission from accessibility; select another source or ask the relevant rights holder for permission.
A license with broad reuse permissions can still carry obligations. The “Creative Commons Attribution 4.0 International Deed” says that CC BY 4.0 allows sharing and adaptation, including for commercial purposes, subject to its conditions, including appropriate credit, a link to the license, and an indication of changes. That describes what the deed permits when that license governs the work; it does not establish that a particular item is CC BY 4.0, that the person posting it controls every element, or that every intended use satisfies the conditions. Record the attribution you will provide and the edits you make, but verify the notice on the actual item rather than borrowing a license label from a similar page.
A compact register can be a table in your project notes: “item and creator; URL; license or terms and version; planned changes; audience; AI processing; attribution or other condition to meet.” Keep a copy or dated record of the notice you relied on, since pages and terms may later change. If the source has no clear terms, the terms prohibit a planned adaptation or use, or you cannot tell whether the license covers the item, the cleanest decision is to replace it with material whose provenance and conditions you can explain, or to obtain permission scoped to your project. Neither a permissive-looking label nor public access settles rights in every element bundled into a page.
Sources: Creative Commons Licenses; Creative Commons Attribution 4.0 International Deed
Will a safe substitute preserve the real work?
A substitute preserves a task only when it keeps the judgment that determines the answer, not merely the visible sequence of steps. Before choosing an analogue, write down four things: the consequential decision being made; the evidence needed to make it; the ambiguity or unusual condition that could change the answer; and the cost of getting the answer wrong. Then compare the substitute against each item. This is a fidelity test derived from the task and context questions in NIST’s AI RMF Core, not a validated score or proof that a portfolio example represents a whole job.
The first item is the decision, stated at the level where a person could act on it. ‘Extract the date from a document’ may be a narrow task; ‘decide whether this request qualifies for an exception’ may depend on dates, missing records, policy context, and consequences. A public or independently authored case can be useful for demonstrating extraction if extraction is the skill being shown. It may be a poor analogue for exception handling if the source records omit why the exception exists or what evidence is authoritative. Similar-looking fields do not guarantee similar judgment.
Next list the evidence and uncertainty that actually bear on the decision. Does the example include conflicting information, incomplete inputs, an unusual but consequential case, or a point where the matter should be escalated? If real work depends on knowing that a record is stale, that two sources disagree, or that a human must pause before acting, a tidy sample with one obvious answer has removed part of the task. The omission matters even when the substitute is well written: it can show that a tool handles clean inputs while saying little about performance under the conditions that call for expertise.
Consider a clearly hypothetical contrast. An independently authored set of sample forms might let a portfolio project show whether a system can extract requested fields and flag a blank field for review. A separate task might involve deciding whether a rare exception is supported by a chain of evidence that is incomplete or inconsistent. Replacing that chain with a neat public example would make the demonstration easier to share, but if the missing context is what drives the decision, the project would no longer test the same judgment. The responsible adjustment is to narrow the claim to extraction, or choose another task whose meaningful reasoning can be demonstrated safely.
NIST’s AI RMF Core treats intended purpose, use context, system limits, measurement, and human oversight as relevant things to document. That structure helps expose where a replica differs from the work it is meant to illustrate; it does not certify fidelity. Use the comparison to state what the case does and does not contain. For example, report that the sample tests field extraction on complete forms, while exception evidence and escalation decisions are outside its scope. That is more informative than calling it a realistic simulation without showing which conditions make it realistic.
Privacy and task fidelity are separate checks. The UK Information Commissioner’s Office guidance “Anonymisation” explains that removing direct identifiers alone may not prevent identification when information can be linked with other material. A faithful-looking redaction can therefore leave contextual clues, while removing enough detail to reduce that risk can also erase the evidence needed for sound judgment. The ICO page is UK personal-data guidance and is under review; it does not settle confidentiality, intellectual property, contract terms, or permission to use material. Passing one check does not answer the other.
A useful substitute may be less role-specific than the original, and that is a limitation to disclose rather than a flaw to hide. If it retains the decision structure, evidence types, meaningful ambiguity, and escalation point, it may demonstrate a bounded transferable skill. If it omits the very condition that changes the decision, describe the narrower operation or choose a different task. If the relevant context cannot be safely recreated, do not fill the gap with a polished replica that invites a broader conclusion than its cases support.
Sources: AI RMF Core; Anonymisation
What does synthetic data demonstrate—and what can it distort?
Synthetic data are generated from a source dataset or model rather than independently authored examples. That difference matters: a fictional case written from general knowledge does not inherit patterns from an employer’s records in the same way that data generated from those records may. Synthetic examples can be useful when there is a defined purpose and permission for the full process, but the label does not establish that they are private, representative, or suitable for a public portfolio.
Use employer-derived synthetic data only when the relevant owner has authorized the source data, generation method, tools or services involved, resulting files, intended audience, and any storage or onward sharing. Authorization for one step should not be assumed to cover the others. If those conditions are unclear, an independently authored scenario or a different task is a cleaner default. This is a practical handling rule, not legal advice; applicable obligations depend on the material and the people or organizations that control it.
If use is authorized, decide in advance what properties the generated examples are meant to preserve. These might include broad field relationships, common decision branches, or a particular pattern of missing evidence. Compare the generated cases with reference properties that are themselves permitted to inspect, and name which comparisons you made. If you did not check rare exceptions, subgroup behavior, or linkage risk, say so. A close match on a selected summary does not establish fidelity on unexamined cases, and a small portfolio sample cannot settle every property of the source material.
NIST’s “Guidelines for Evaluating Differential Privacy Guarantees (NIST SP 800-226)” discusses utility challenges that can arise with synthetic data. It says that, in some cases, accuracy for subpopulations can fall; added uncertainty can propagate; and generative bias can propagate. These are possible challenges, not a finding that every synthetic dataset has those problems or that a particular employer-derived sample has them. The report concerns privacy guarantees and utility, not whether a portfolio project accurately represents a specific workplace task.
The practical implication is to avoid treating an average-level resemblance as proof that difficult cases survived generation. A generated set may preserve selected common patterns yet contain few examples of rare conditions, simply because those conditions were scarce or were not a generation target. That is a reason to ask what was represented and checked, not a result about any particular dataset. Where rare cases drive consequential decisions, an untested sample should not be presented as evidence that the workflow handles them; narrow the claim or use a permitted evaluation method that addresses them.
The UK government’s guidance “Ethical considerations relating to the creation and use of synthetic data” frames assessment around purpose, data quality, privacy, and potential harms across creation and use. It supports asking those questions at each stage; it does not validate a dataset, guarantee representativeness, or authorize use of employer records. In a portfolio description, distinguish the origin of the material, the properties you intended to retain, the checks you actually performed, and the limitations that remain. Those are separate statements, and none should imply a broader privacy guarantee than the method supports.
Do not claim differential privacy merely because records were altered, names were removed, or a generator produced new rows. Differential privacy is a specific privacy framework with guarantees that depend on a defined method and its evaluated parameters; SP 800-226 is guidance for examining such guarantees, not a certificate for a dataset. Unless the process and guarantee have actually been assessed, describe the material plainly as synthetic and explain the generation and evaluation limits that are permitted to disclose. If that explanation itself would reveal restricted details, omit the data and demonstrate the skill with an independent case instead.
So synthetic data can be a considered option when its source and creation are authorized and its intended utility is checked against permitted references. It cannot, by itself, prove confidentiality, preserve every rare decision condition, or establish workplace performance. Keep the conclusion proportional to the evidence: identify the selected patterns, checks, and gaps, and limit the project claim to those cases and properties. When permission or fidelity remains unresolved, changing the task is more defensible than treating ‘synthetic’ as a clearance certificate.
Sources: Guidelines for Evaluating Differential Privacy Guarantees (NIST SP 800-226); Ethical considerations relating to the creation and use of synthetic data
How can I test one bounded AI step honestly?
A modest evaluation is enough to make a small demonstration inspectable if it defines the operation, the cases, the expected behavior, and how a person checked the result. Begin with one task and one AI step: for example, have a system propose a category for a request, rather than claim to automate the whole process of deciding what should happen to it. Write down in advance what a useful answer should contain, what should trigger a flag or a human decision, and what the system should do when evidence is missing. This makes the test about an observable operation instead of a broad claim about intelligence or workplace performance.
Build a small set of cases from independently authored examples or material whose use is permitted for the project and its AI processing. Choose cases to probe the conditions that matter to the operation, not just a row of easy successes: ordinary inputs, ambiguous wording, missing information, conflicting evidence, and relevant edge cases. The task determines which variations matter. For a field extraction step, a blank or differently formatted field may be informative; for a proposed routing category, two plausible categories or insufficient evidence may matter. These are candidate test conditions, not a claim that any particular model will fail them. If the input is synthetic or simulated, say so and make clear what that choice permits the project to test.
For each case, record the expected outcome and why, the AI output, the check you applied, and the reviewer’s decision: accept, correct, reject, or escalate, as appropriate. Preserve enough context for a reader to understand the check without exposing restricted source material. Note any change you made to the prompt, rule, or output before reaching the final result; otherwise the reviewer cannot tell whether the displayed answer came from the model or from human repair. Include a limitation beside the case when the test omits information or conditions that could change the decision. The evaluation can be a simple table rather than a benchmark suite.
If you report counts, define the unit and denominator. ‘Eight of ten cases matched the prewritten expected result’ is interpretable only if the reader can tell what counted as a case, what counted as a match, how the ten cases were selected, and who judged ambiguous results. A percentage from a tiny, hand-selected set can imply statistical precision the design does not support. It may be more useful to show the cases and decisions directly, then state that the sample is exploratory and limited to those examples. Do not invent a miss to make the table look rigorous, and do not hide an actual miss because it complicates the story; report only what the record shows.
NIST’s “AI RMF Core” supports documenting context, measurement, oversight, and system limitations, while “Artificial Intelligence Risk Management Framework (AI RMF 1.0)” states that the framework is intended to be voluntary. It is a structuring aid for asking and recording risk questions, not certification, a required portfolio checklist, or permission to use source material. A small test designed this way can show how the operation behaved on the cases presented and what human review contributed. It cannot on its own establish accuracy across a population, productivity gains, production reliability, hiring value, or job readiness. Those conclusions need other evidence and a different scope.
There is a real tradeoff: a small case set may be too narrow or subjective to support a percentage, while omitting evaluation leaves a viewer unable to see what was checked at all. The useful middle ground is transparent scope. Describe how the cases were made, the expected behavior, the review rule, the observed outputs, and what was left untested. If another person would reasonably judge an ambiguous case differently, label the judgment and explain the criterion rather than presenting it as an objective model score. That makes the demonstration more inspectable without pretending a small exercise settles how the workflow would behave on unfamiliar work.
Sources: AI RMF Core; Artificial Intelligence Risk Management Framework (AI RMF 1.0)
What should a reviewer be able to inspect?
Give a reviewer a short route through the project: what the task is, what material it uses, what the AI contributes, how a person checks the result, and where the demonstration stops. A concise README can orient someone who opens a repository; GitHub’s “About READMEs” describes it as information visitors see when viewing that repository. That establishes a useful navigation role, not a hiring advantage. Put the main boundary near the top, then link to the example and the evidence table so the reader does not have to infer the project’s purpose from filenames or interface polish.
The minimum useful package can be small: a short README; one viewable or runnable example; source, license, and permission notes; a compact case-and-result table; a statement of what the AI did; and a note on what the person checked, changed, or decided. Add what remains untested, including important conditions omitted from the sample. If inputs cannot be shown, describe their origin and permitted scope without implying the hidden material proves fidelity. Label independently authored, public, synthetic, simulated, and generated artifacts accurately. A single public repository, “AI portfolio example with evidence labels and explicit project boundaries,” illustrates one way to mark such categories and limits; it is a presentation example, not evidence that this format is universally effective.
Separate implementation artifacts from evaluation evidence. Code, prompts, configuration, and a running example let a reviewer inspect how the operation was assembled and what it is intended to do. The case table records what happened on the cases actually tried, including the human check and decision. One cannot substitute for the other: readable code does not show that outputs were checked, and a results table without enough method or provenance does not show what produced those results. If the project includes a generated dataset or simulated workflow, state how it was created at a level safe to disclose and which properties or behaviors the displayed cases do and do not demonstrate. A useful row can identify the case, the expected result, the output shown, the verification performed, and the final human decision; add a short note when the expected result depends on an explicit judgment rule. Include a link or filename for the relevant code or example when that connection is not obvious. This lets the reviewer move from claim to implementation to observed case without confusing an attractive interface with an evaluation record. Keep that trail concise and easy to follow.
A custom interface is optional. It earns its place if it helps a reviewer follow the input, AI contribution, check, and decision; it adds little if the same sequence is clearer in a short table or notebook. GitHub’s “Using your GitHub profile to enhance your resume” recommends adding context to repositories shown on a profile or resume, but that is platform advice, not evidence about recruiter behavior. Treat presentation as a way to reduce navigation friction. Do not treat visual finish, a polished demo, or a repository’s existence as proof of accuracy, workplace impact, hiring value, or readiness for a role.
Sources: About READMEs; Using your GitHub profile to enhance your resume; AI portfolio example with evidence labels and explicit project boundaries
When should I change tasks or ask for private-demo approval?
If the decision depends on context you cannot safely reproduce for public viewing, choose between a different task and a specifically scoped private demonstration. Start by asking what the project needs to prove. If a permitted, independently authored or public task can still show the relevant reasoning, switching may be the simpler route. It can cost time: you may need to define a new case, adjust the AI step, and explain why the substitute demonstrates a narrower skill. That cost is worth comparing with the effort and uncertainty of getting approval; neither route automatically preserves the original task’s full context. A useful comparison is the decision-relevant context gained against the extra coordination, restrictions, and explanation each route requires.
A private demo can retain context that a public example would lose, but private access is not permission by itself. Redaction and access controls may reduce exposure, yet they do not establish that the material may be used, processed by a particular tool, shown to particular people, or retained. Ask the relevant data, employer, or client owner for explicit scope before building around restricted inputs. Make the request concrete: identify the source material and intended AI tools or processing; who may view the demonstration; whether it may be recorded or stored and where; how long access lasts; and whether anyone may copy, forward, or otherwise share it. Include the intended purpose and what the project will show so the owner can assess the actual request. Do not assume a verbal or informal yes to one part covers the rest.
Then compare the likely value of that access with its conditions. If the owner permits a live, named-audience demonstration but prohibits recording or onward sharing, plan a demo that can meet those limits and describe them accurately. Do not later turn it into a public repository, screenshots, or a recording without separate permission. If approval would require more coordination than the project can support, or the allowed audience would not serve your purpose, a permitted substitute may be more practical. Preserve the distinction between showing a bounded skill and claiming to have reproduced the full work context.
If neither a faithful permitted substitute nor scoped approval is available, do not publish the material. You can still describe the independent skill demonstration: the decision structure you designed, the permitted inputs used, the AI step, and the human checks. State which workplace context or judgment it does not establish. This is a practical decision process, not legal counsel; the person or organization controlling the material may have requirements you cannot infer from its appearance or from technical safeguards.
What can I finish next within my constraints?
Choose one recurring task, name the judgment it requires, and identify one permitted source route. Before coding, write a one-paragraph project brief stating the bounded AI step, the cases you can use, the human check, and the claim the demonstration will support; add a short source and permission note. If no permitted source preserves the judgment you want to show, switch tasks or request scoped approval first. The free task checker is optional if you also want to examine broader task-level change pressure; it is not a portfolio selector or a personal probability of job loss. A paid career roadmap can compare broader scenarios and constraints, but it does not guarantee employment or income. Keep the brief small enough to finish around your available time, access, and responsibilities: one task and a few deliberately chosen cases can make a bounded demonstration inspectable without pretending to represent every condition. The immediate decision is whether your proposed source is permitted and whether it retains the judgment your project is meant to show. If either answer is unresolved, settle that question before investing in implementation or presentation.
Sources and notes
- Guidelines for Evaluating Differential Privacy Guarantees (NIST SP 800-226)
Synthetic data may introduce utility challenges: in some cases accuracy for subpopulations can fall, added uncertainty can propagate, and generative bias can propagate; these are possible challenges, not inevitable outcomes.
- Artificial Intelligence Risk Management Framework (AI RMF 1.0)
NIST states AI RMF 1.0 is intended to be voluntary, rights-preserving, non-sector specific, and use-case agnostic.
- AI RMF Core
The Core organizes desired outcomes and actions across Govern, Map, Measure, and Manage, including documenting context, measurement, oversight, and limitations; it does not grant permission to disclose work material.
- Anonymisation
Anonymisation concerns whether people can be identified, including through linkage; removing direct identifiers alone is not a complete identity-risk test.
- Ethical considerations relating to the creation and use of synthetic data
Synthetic data require consideration of purpose, data quality, privacy and potential harms across creation and use; guidance does not authorize disclosure or establish representativeness for a specific portfolio task.
- Creative Commons Licenses
Public availability does not erase reuse conditions: CC licenses vary in attribution, commercial-use, adaptation and share-alike requirements.
- About READMEs
GitHub describes README files as information visitors see when viewing a repository; it supports orientation, not hiring outcomes.
- Using your GitHub profile to enhance your resume
GitHub suggests adding context when showcasing repositories on a profile or resume; this is platform advice and does not demonstrate recruiter response or employment effects.
- AI portfolio example with evidence labels and explicit project boundaries
One self-published repository labels public-source material, synthetic data, simulation and generated artifacts and states project limits; it illustrates a presentation choice only.
- Creative Commons Attribution 4.0 International Deed
CC BY permits sharing and adaptation, including commercial use, subject to attribution and other specified conditions; a license must be read at the item level.
Apply this to your own work
See the whole job market at once.
Explore which occupations AI may reshape, then turn the signal into a practical response.
Explore the job map