Recruiting, consent and research ethics

A new researcher at MealBox arrived with a template that looked professional: a consent sheet labelled “Agreement”, an advert on an online research panel offering 500 rupees to “anyone interested in food”, and an incentive paid only if the feedback was “useful”. By the time the first session started, three things had gone wrong. The participant had never ordered a lunch service in their life; the sheet never said how long the recording would be kept; and nobody had asked whether the person knew they could stop.

Each of those pieces of the recruit sits on a page of the GOV.UK Service Manual, and the costs of getting them wrong fall on participants, not the team. This lesson is the small, specific list of things that have to be right: the screener that filters by behaviour, the sheet that makes consent informed, the data plan that protects what people shared, and the incentive and support that make the whole arrangement fair.

Recruit for what people did, not only who they are

Demographic recruiting — “women aged 25 to 40 in Pune” — looks precise and tells you very little about who will show up. The GOV.UK Service Manual’s page on finding user research participants puts the alternative plainly: recruit “based on behaviour and attitudes rather than demographics”. A screener that asks “in the last four weeks, did you order lunch delivery at work?” filters for people whose past behaviour matches the problem; a demographic filter produces a room full of people who may or may not have ever used a service like yours.

Three practical rules keep a screener honest. First, add decoy answers so the “right” option is not obvious; a behavioural question with four options, only one of which qualifies, invites participants to notice which answer gets them in. Second, include people who do not use digital services when the research is about them; the Service Manual’s discovery guidance is firm that research must include people with low digital skills, which no online panel reaches on its own. Third, write what the research is about and what will happen in plain language on the recruit advert itself, so people can decide before they fill the screener in.

Five channels do most recruiting between them, each with a different cost in money, time and reach:

For the MealBox cancellation study, the right mix is usually an email to recent sign-ups for most of the sample and outreach through a local accessibility group for one or two participants who use assistive technology. A panel is a poor fit on its own, because panels skew digital and salaried; it is useful as a quick first pass when the research is about a well-covered task.

Informed consent is a conversation, not a signature

The signature is the smallest part of consent. For a participant’s agreement to be informed, the GOV.UK Service Manual’s page on getting consent lists what they must understand before they agree: who is doing the research, the purpose, what data you collect, what will happen during the research, how you will use the results and who you share them with, that participation is voluntary and that they can stop or withdraw at any time, how long their data will be kept, and what their rights are and how to complain. The sheet also has to say whether and how the session is recorded, and who else is watching.

A one-page information sheet carries all of those items and lands with the participant before the session. The Service Manual recommends sending it during recruitment so people have time to read it, ask questions and arrive prepared. On the day, consent is captured at the start of the recording, item by item, with the researcher reading the sheet aloud so verbal consent is as valid as a signature. The five moments where the conversation goes right or wrong fit on one scene:

Two sentences in that walkthrough do most of the work. The first is “you can stop any time, without losing the incentive”, which the manual asks be clear during recruitment and reminded again at the start and end of the session. The second is the deletion commitment: when a participant withdraws, the recording and the notes about that participant are deleted, and the incentive is still paid. A design critique that does not treat withdrawal as a normal response has already got consent slightly wrong, because the pressure to carry on is already in the room.

Collect less, store it well, delete it on schedule

Everything a round of research records is personal data, and the GOV.UK Service Manual’s page on managing user research data treats it that way: notes, audio, video, screenshots, questionnaire responses and the paperwork participants refer to. Three habits make the difference between a safe round and one that has to be apologised for later.

Collect only what you need. If audio answers the research questions, do not record video; if video is needed for one task, delete the rest after transcription. The habit saves time at analysis and keeps the privacy cost on participants small. The example here checks a MealBox consent sheet against the eight required elements, and prints the draft’s missing or weak fields next to a corrected version:

"""A MealBox consent sheet, scored against GOV.UK's informed-consent list.

Every participant, incentive and sentence below is invented for this lesson.
The eight required elements come from the GOV.UK Service Manual's page on
getting informed consent for user research, and the data-handling section
from its page on managing user research data. Lesson pins Python 3.14.8.
"""

# A flawed consent sheet, as it was first drafted
DRAFT = {
    "who": "A research team",                              # should name the organisation (data controller)
    "purpose": "To make the app better",                   # vague; needs the actual research question
    "data_collected": "Audio, video, notes, screen",       # ok
    "who_sees": "The team",                                # ok, but needs to name third parties
    "use_results": "Internal use",                         # ok
    "voluntary": "You can stop any time",                  # ok, but should name "without losing the incentive"
    "retention": "",                                       # MISSING
    "withdraw": "Email us if you want to withdraw",        # ok
    "recording": "We will record the session",             # should say who sees and how long it is kept
    "complaint": "",                                       # MISSING: how they can complain
    "incentive": "Paid at the end if feedback is useful",  # TIED to positive feedback (not allowed)
    "access_need": "",                                     # MISSING
}

REQUIRED = {
    "who": {"must_name": ["mealbox"], "note": "name the organisation responsible for the data (data controller)"},
    "purpose": {"must_name": ["cancel", "sign-up", "pause"], "note": "state the research question, not a general aim"},
    "data_collected": {"must_name": [], "note": "list each kind of data (audio, video, notes, screen)"},
    "who_sees": {"must_name": ["team", "transcription", "third"], "note": "name the team and any third parties"},
    "use_results": {"must_name": [], "note": "state the use and who the findings are shared with"},
    "voluntary": {"must_name": ["stop", "withdraw", "without losing"], "note": "state that stopping does not cost the incentive"},
    "retention": {"must_name": ["year", "months", "delete"], "note": "state the retention period and when the data is deleted"},
    "withdraw": {"must_name": ["email", "delete"], "note": "describe how to withdraw and what happens to the data"},
    "recording": {"must_name": ["recording", "stored", "kept"], "note": "state what is recorded, who sees it and how long it is kept"},
    "complaint": {"must_name": ["complaint", "contact", "email"], "note": "state how to raise a complaint or ask about their rights"},
    "incentive": {"must_name": ["paid", "regardless", "whether", "any outcome"], "note": "pay regardless of what participants say (never tied to 'useful' or 'helpful')"},
    "access_need": {"must_name": ["bring", "assistive", "support", "format", "sign"], "note": "ask about access needs and how to meet them"},
}


def score(sheet):
    """Return (findings, missing). findings is a list of (field, note); missing is a count."""
    findings = []
    for field, rule in REQUIRED.items():
        text = (sheet.get(field) or "").lower()
        if not text:
            findings.append((field, "missing"))
            continue
        if rule["must_name"] and not any(word in text for word in rule["must_name"]):
            findings.append((field, rule["note"]))
    # Specific check: incentive tied to "useful" or "helpful" feedback
    inc = (sheet.get("incentive") or "").lower()
    if any(word in inc for word in ("useful", "helpful", "positive")):
        findings.append(("incentive", "incentive is tied to the kind of feedback given, never allowed"))
    return findings


# Draft fails
draft_findings = score(DRAFT)
print(f"Draft consent sheet: {len(REQUIRED) - len(draft_findings)} of {len(REQUIRED)} elements pass")
for field, note in draft_findings:
    print(f"  [{field:<13}] {note}")

# Corrected sheet, following the lesson's list
FIXED = dict(DRAFT)
FIXED.update({
    "who": "MealBox, a lunch-subscription service in Pune, is the data controller",
    "purpose": "To learn why new subscribers cancel in their first week and improve the pause flow",
    "who_sees": "The MealBox research team and one third-party transcription service",
    "voluntary": "You can stop any time, without losing the incentive",
    "retention": "Audio and video for 12 months, notes with ids for 2 years, then both are deleted",
    "withdraw": "Email research@mealbox.example; your data will be deleted within 30 days",
    "recording": "Audio and video of the session, stored securely, kept 12 months then deleted",
    "complaint": "Any complaint: email research@mealbox.example or write to the postal address",
    "incentive": "1,200 rupees by UPI the same day, paid regardless of what you say or whether you finish",
    "access_need": "Tell us in advance about any assistive technology or support you would like us to bring",
})

print()
fixed_findings = score(FIXED)
print(f"Corrected sheet: {len(REQUIRED) - len(fixed_findings)} of {len(REQUIRED)} elements pass")
for field, note in fixed_findings:
    print(f"  [{field:<13}] {note}")

Draft consent sheet: 5 of 12 elements pass
  [who          ] name the organisation responsible for the data (data controller)
  [purpose      ] state the research question, not a general aim
  [retention    ] missing
  [recording    ] state what is recorded, who sees it and how long it is kept
  [complaint    ] missing
  [access_need  ] missing
  [incentive    ] incentive is tied to the kind of feedback given, never allowed

Corrected sheet: 12 of 12 elements pass

Store it securely from the start. Move recordings off cameras and phones into secure storage as soon as the session ends. Pseudonymise transcripts before they are shared more widely, which means replacing names, addresses and account numbers with participant ids. MySmartCoPilot’s own text redactor and PII redactor run in the browser, with the raw text never leaving the device, which fits the “collect less, share less” rule.

Delete on a named schedule. The Service Manual suggests two or three years as a reasonable retention period for user research data, with sooner if the data is no longer needed. Name the period on the information sheet, write it in the plan, and set a calendar reminder for the deletion.

Research with disabled participants takes planning, not permission

Research that leaves disabled participants out is research the design will fail on, so plan sessions that include them from the start. The Service Manual’s page on running sessions with disabled people gives a practical set of steps, which fall into three groups.

Ask before you plan. The recruit asks about assistive technology, support needs and preferred format for the information sheet. If a participant uses a screen reader, send the sheet as a text file or an HTML page they can read with it; if a participant has a motor impairment, ask whether a verbal consent on the recording is easier than signing a form. For sessions with people who are deaf or hard of hearing, book a sign-language interpreter and use a video call tool whose captions they trust.

Plan the session around the person. Allow extra time for consent and set-up, especially if the participant brings their own technology; their personal settings are hard to recreate on another device, so bringing their own set-up means the session measures the task rather than the setup. Where this is not possible, visit the participant instead. Allow breaks, pay for preparation time where it is substantial, and avoid lab venues when the research could happen in a familiar place.

Behave in the room. The Service Manual asks researchers to talk directly to the participant rather than to any interpreter, to use everyday language without worrying about causing offence, and to ask instead of guessing. “See you later” is fine; guessing what somebody said is not. If you are unsure which terminology to use, ask the participant what they prefer.

A small test for every plan: can a person using the technology in your recruit actually read the information sheet, sign the consent form (or confirm it another way), and do the session on their own set-up? If not, the plan excludes them before it starts.

Fair incentives, and care with sensitive topics

Fair pay beats a clever incentive. The incentive is for time, which means every participant doing the same work for the same length gets the same amount, in a form they can use. Keep the amount on the information sheet, pay promptly, and never tie the payment to the content of the feedback. “Paid if your feedback is useful” and “paid after a review” both put the money at the researcher’s discretion, which is where unspoken pressure comes from.

Sensitive topics need preparation. The Service Manual’s page on researching emotionally sensitive subjects — on money, health, bereavement, abuse — asks for three habits. Prepare signposting: a short list of support services you can share if a conversation touches a difficult moment. Allow pauses: tell participants in the recruit and again at the start that stopping is normal, and watch for signs that the conversation is costing more than you expected. Never push: if someone goes quiet or steers away, go with them, and leave the question for later or for another method entirely.

Children and vulnerable adults need the consent of a parent, guardian or carer as well as their own agreement, and both the questions and the setting benefit from the input of someone with experience in that work. If the project is new territory for the team, invite a specialist before writing the first screener. The lesson’s own cases, MealBox and Ward Desk, deliberately stay away from research with children because that specialist step happens outside the lesson.

The next lesson, interviews and field research, takes the plan and the ethics into the session itself. The lesson after it, surveys and behavioural data, picks up the counting part of the plan.

Practise

A flawed MealBox consent sheet, with a screener to match: five fields need the sentence the Service Manual asks for, and two decoy answers plus an access-needs field make the screener honest. Open the file and fill in each field:

python -m pytest test_sheet.py

"""Starter: fill in the TODOs so the consent sheet and screener survive the tests in test_sheet.py."""

SHEET = {
    "data_controller": "",                                              # TODO: name MealBox as the responsible organisation
    "research_question": "To make the app better",                       # TODO: state the real question
    "data_collected": "Audio, video, notes, screen recording",
    "data_shared_with": "The research team",                             # TODO: name any third parties if they will see the data
    "voluntariness": "Participation is voluntary",                       # TODO: add that stopping does not change the incentive
    "retention": "",                                                     # TODO: state a period and the deletion
    "withdrawal": "Email research@mealbox.example to withdraw",          # TODO: say what happens to the data on request
    "recording": "The session will be recorded",                         # TODO: name what is recorded, who sees it and how long
    "complaint": "",                                                     # TODO: give a contact for complaints
    "incentive": "A payment at the end if your feedback is helpful",     # TODO: rewrite so payment is unconditional
    "access_need": "",                                                   # TODO: ask about assistive technology and extra time
}

SCREENER = {
    "demographic": "What is your age range?",
    "behavioural": "",                     # TODO: ask about something the participant has done, with at least four options
    "decoys": [],                          # TODO: list the decoy answers in the behavioural question (at least 1)
    "access_language": "",                 # TODO: free text for assistive technology and preferred language
}

from sheet import SCREENER, SHEET


def _lower(key):
    return (SHEET.get(key) or "").lower()


def test_data_controller_names_mealbox():
    assert "mealbox" in _lower("data_controller"), "name MealBox as the data controller"


def test_research_question_is_specific():
    text = _lower("research_question")
    assert any(word in text for word in ("cancel", "pause", "subscribe", "sign-up", "sign up")), \
        "state the actual research question, not a general aim"


def test_voluntariness_names_the_incentive():
    text = _lower("voluntariness")
    assert "stop" in text or "withdraw" in text, "name the right to stop"
    assert "incentive" in text or "payment" in text or "paid" in text, "state that stopping does not change the incentive"


def test_retention_has_a_period_and_deletion():
    text = _lower("retention")
    assert any(unit in text for unit in ("month", "year", "day")), "name a retention period"
    assert any(word in text for word in ("delete", "deleted", "destroy", "removed")), "state when the data is deleted"


def test_withdrawal_says_what_happens_to_the_data():
    text = _lower("withdrawal")
    assert any(word in text for word in ("delete", "deleted", "removed", "destroyed")), \
        "say what happens to the data on withdrawal"


def test_recording_names_what_who_how_long():
    text = _lower("recording")
    assert any(word in text for word in ("audio", "video", "screen")), "name what is recorded"
    assert any(word in text for word in ("team", "transcription", "controller", "stored")), "name who sees the recording"
    assert any(unit in text for unit in ("month", "year", "day")), "name how long it is kept"


def test_complaint_has_contact_details():
    text = _lower("complaint")
    assert "@" in SHEET.get("complaint", "") or "phone" in text or "write" in text, \
        "give an email, phone or postal address for complaints"


def test_incentive_is_paid_regardless():
    text = _lower("incentive")
    assert any(word in text for word in ("regardless", "whatever", "unconditional")), \
        "state that payment does not depend on what participants say"
    for word in ("helpful", "useful", "positive"):
        assert word not in text, f"never tie payment to '{word}' feedback"


def test_access_need_offers_support():
    text = _lower("access_need")
    assert any(word in text for word in ("assistive", "support", "access", "sign language", "screen reader", "braille")), \
        "ask about assistive technology or support"
    assert any(word in text for word in ("time", "format", "advance")), "offer extra time or an alternative format"


def test_data_shared_includes_third_parties():
    """The sheet mentions the team and at least one third party, if any work with the data."""
    text = _lower("data_shared_with")
    assert "team" in text or "mealbox" in text, "name the research team"
    assert any(word in text for word in ("transcription", "third", "none", "no third", "only the")), \
        "state whether any third party will see the data"


def test_screener_has_a_behavioural_question_with_decoys():
    text = SCREENER["behavioural"].lower()
    assert any(verb in text for verb in ("have you", "did you", "last time", "in the last")), \
        "ask about something the participant has done"
    assert len(SCREENER["decoys"]) >= 1, "add at least one decoy answer"


def test_screener_covers_access_and_language():
    text = SCREENER["access_language"].lower()
    assert any(word in text for word in ("assistive", "support", "screen reader", "sign language", "braille")), \
        "mention assistive technology"
    assert any(word in text for word in ("language", "hindi", "marathi", "english", "preferred")), \
        "mention preferred language"

The self-check that follows pairs the eight consent items with the recruiting rules:

Interview questions

Warm-up (fresher to 2–5 years): what makes consent informed?

Consent is informed when a participant understands, before they agree, who is doing the research and why, what will be recorded, who sees it, how long it is kept and how they can stop. GOV.UK’s Service Manual puts the eight items on one page: who, purpose, data, process, use, voluntariness, retention and complaint route. A signature on its own proves none of that; the sheet that lands before the session carries it, and the researcher confirms each item at the start of the recording. Two sentences do most of the work: “you can stop any time without losing the incentive”, and the commitment that on withdrawal the data is deleted. Treat consent as a conversation that starts during recruitment, not a form signed in the first minute of the session.

Two harder questions follow: one designs an incentive with no pressure in it, the other handles discomfort that surfaces mid-session. Write your own first move down, then compare it with the step published beside each:

Key takeaways

• A screener filters by past behaviour with decoy answers, so the “right” option is not obvious; demographics alone do not say who has used the service you are studying.
• Informed consent names eight things on a one-page sheet before the session, repeats them at the start of the recording, and treats withdrawal as a normal response the incentive does not depend on.
• Research data is personal data: collect only what the questions need, store it securely from the first minute, and name a retention period of two or three years with a scheduled deletion.
• Sessions with disabled participants are planned around the person’s own assistive technology, with extra time for consent and set-up and a visit when the technology cannot travel.
• Pay fairly and promptly for time, in a form the participant can use, and never tie payment to the content of the feedback or to finishing every task.