AI Survey 2026  ·  L4WB Group

How is AI used
at L4WB?

In April 2026, 40 of 42 colleagues across the L4WB Group (95%) completed a survey on how they use artificial intelligence, what concerns them, and what they need from an AI policy. This is what they said.

Learning for Well-Being Institute

The Current Picture

AI is already part
of how we work

  • 90% of colleagues use AI at least a few times a week; 45.0% use it daily.
  • Three tasks dominate: editing, translating, and drafting (75%, 72.5%, and 70% respectively).
  • Usage extends well beyond writing: image generation (37.5%), literature searches (32.5%), data analysis (30.0%), and code (15.0%).

Q3How often do you use AI for work?

47.5%
A few times a week
19 of 40 colleagues
45.0%
Daily
18 of 40 colleagues
7.5%
A few times a month
3 of 40 colleagues

Q1What do you use AI for? (% of 40 respondents)

Editing, proofreading, rephrasing
30
75%
Translating text
29
72.5%
Drafting written content
28
70%
Summarising documents
20
50%
Brainstorming / planning
20
50%
Researching topics
18
45%
Literature searches
13
32.5%
Data analysis
12
30%
Generating images / media
15
37.5%
Writing / debugging code
6
15%

% of 40 respondents. Respondents could select multiple tasks.

Q2Which tools?

ChatGPT leads, with four in five colleagues (82.5%) using it. Gemini (45.0%), Claude and Copilot (25.0% each) follow at some distance. Most colleagues use more than one tool. A handful use L4WB’s own AI tools: LitSearch and AutoMap.

Six Scenarios · Where Colleagues Stand

Six scenarios,
two clear dividing lines

The survey presented six real-world AI use cases and asked colleagues to rate each: acceptable, acceptable with caveats, or not acceptable. Four produced broad agreement. Two produced genuinely split responses; these are the ones the AI policy needs to address directly.

Scenarios are ordered from most broadly accepted (top) to clearest “no” (bottom). The two with the most divided responses (flagged) show where colleagues hold different views and where a clear policy position is needed.

Key:
Yes, this is fine
Yes, but with caveats
No, this shouldn’t be done
Not sure
S1: Brainstorm project framing with ChatGPT
18%
62.5%
S2: AI summarises a published academic paper
32.5%
57.5%
S6: AI writes first draft of proposal rationale ▶ most divided
30%
45%
18%
S4: AI translation sent directly to funder ▶ policy clarity needed
65%
32.5%
S5: Lit review claims from AI summaries only
70%
S3: Confidential partner draft pasted into ChatGPT
82.5%

n=40 respondents. Hover over segments to see precise counts.

Most divided responses

AI writes first draft of proposal rationale

You're drafting a project proposal and use an AI tool to write the first draft of the context and rationale section, which you then edit and refine.

Yes, fine
30% (12)
With caveats
45% (18)
Should not be done
17.5% (7)
Not sure
7.5% (3)

The policy needs a clear position, with a worked example.

Policy clarity needed

AI translation sent directly to a funder

You've written a report that will go to a funder. You use an AI tool to translate it into Spanish and send the translated version directly to the funder.

Yes, fine
2.5% (1)
With caveats
65% (26)
Should not be done
32.5% (13)

The key question is not whether to use AI for translation, but whether it has been properly reviewed before sending.

Where we are clear

  • Sharing a confidential partner report with a public AI tool: 82.5% say no.
  • Pulling claims from AI summaries without reading original papers: 70.0% say no.

The policy needs to confirm these as firm boundaries: they reflect what most colleagues already believe.

Risks We Recognise

What concerns
our colleagues

Two concerns were raised equally: accuracy and over-reliance, each cited by 65% of colleagues. They are two sides of the same risk: when output is not checked, both accuracy and skill development suffer.

Data protection and copyright are both at 50%, marking a clear expectation that the policy will set firm boundaries on what can and cannot be shared with AI tools.

Q16Biggest concerns about AI use at L4WB (% of 40 respondents, multi-select)

Accuracy or quality of responses
26
65%
Over-reliance on AI
26
65%
Data protection and privacy
20
50%
Copyright or intellectual property
20
50%
Decrease of technical skills
19
47.5%
Environmental impact
14
35%

Top concerns (multi-select). Accuracy and over-reliance are tied at 65%.

Environmental impact was raised by 35% of respondents, reflecting a genuine awareness of AI’s energy footprint. L4WB’s dedicated Environmental Policy already commits the organisation to integrating environmental considerations across all operations and decision-making, including digital tools and services. AI use sits within that commitment: choosing tools thoughtfully, avoiding unnecessary use, and supporting the Foundation’s broader sustainability goals.

Q17Should staff disclose when AI was used in outputs?

In some cases
24
60%
Rarely
6
15%
Always
5
12.5%
Not sure
4
10%
Never
1
2.5%

60% say "in some cases": disclosure is context-dependent, not a blanket rule.

What this means for the policy

Disclosure is not only about external outputs. Sending AI-generated content to a colleague (a draft, a summary, a translation) also warrants a note. The policy defines when disclosure is required for both internal and external audiences, based on the extent of AI involvement and the context.

Draft Policy · Survey Findings

How the survey aligns
with the draft policy

The survey found broad alignment with the positions already in the draft policy. It also surfaced points that needed clarifying or that the draft had not yet addressed.

Where the survey aligns with the draft policy

On the core questions, survey responses and draft policy point in the same direction. The positions colleagues already hold on data protection, accuracy, and disclosure are the positions the policy takes. This report explains each in context.

Points from the survey not yet in the draft policy

  • Over-reliance on AI tied with accuracy at 65% as the top concern, making it the single most significant finding. The draft does not yet name over-reliance as a principle; the survey makes clear it needs to be.
  • Internal disclosure is a gap: most colleagues think of disclosure in terms of external outputs, but the survey shows that sharing AI-generated work with a colleague also warrants transparency. The draft does not yet cover this.
  • AI-assisted translation for external audiences produced near-unanimous reservations (65% “with caveats”, 32.5% “no”), pointing to a gap the policy needs to fill with a worked example, not just a general position.
  • An enabling tone is not yet explicit in the draft: 70% of respondents asked for practical examples over rules, and 5 of 27 open-ended responses specifically asked for enabling rather than restrictive language. This should shape how the policy is written throughout.

What Colleagues Need

Wants, needs, and
what shaped the policy

Three survey questions asked what colleagues need from a policy. The answers point in the same direction: practical over procedural, examples over rules, enabling over restrictive. Together they shaped specific choices in the draft policy.

Q18What support would help you use AI responsibly? (% of 40, multi-select)

Examples of good practice
28
70%
Written guidance
20
50%
Training sessions
20
50%
Approved tool list
19
47.5%

The top ask: practical examples (70%). Written guidance and training (50% each). Approved tool list (47.5%).

Q19What colleagues haven't tried yet

Several colleagues mentioned tasks they would like to use AI for but held back from, not because they thought it was wrong, but because they were uncertain what was permitted. These include: systematic proposal writing with organisational materials, video production for project communications, and data analysis for non-research roles. Clarity in the policy removes this friction.

Q20Open-ended responses

Themes in written suggestions

27 of 40 responded

One response could cover multiple themes. Counts reflect the number of respondents who raised each theme.

Practical guidance: when and how
9
9 of 27
Human accountability and oversight
8
8 of 27
Data protection and privacy
5
5 of 27
Enabling, not restrictive policy
5
5 of 27
Ethical and responsible use
4
4 of 27
Training and skill development
3
3 of 27
Clear list of approved tools
3
3 of 27
Transparency and disclosure
2
2 of 27

Thematic analysis of Q20 open-ended text responses.

In their own words

Verbatim responses from the open-ended survey question, some condensed for length.

The policy should be enabling rather than restrictive, encouraging thoughtful and widespread use of AI across the organisation. It should provide practical, day-to-day guidance, including common use cases such as translation, meeting summaries, pedagogical writing, content development, and basic design support. At the same time, it should clearly define boundaries around sensitive and confidential information.
The policy should support a balanced and thoughtful use of AI, one that recognises its value as a tool, but keeps human judgement, responsibility, and authentic voice at the centre. AI should not replace our thinking, creativity, or relationships, but rather support them.
I think AI is really a tool that makes me a more efficient worker but I do feel stress sometimes over how I use AI and whether it is aligned with how I should be. Trainings on how we can use AI for improved efficiency would really help.

Survey Into Policy

Six positions that will
shape the AI policy

Based on what colleagues told us, these are the positions the L4WB AI Policy will take. Several were already under consideration in the draft; the survey confirmed, clarified, and in some cases strengthened them.

1
In draft policy

Accountable by default

The staff member owns every output. AI assists; it does not decide, author, or bear accountability. Using AI for a task you cannot verify is not a shortcut. It is a risk.

2
In draft policy

Sensitive data stays protected

Sensitive information about children, families, partner organisations, and beneficiaries does not enter public AI tools. This is a firm boundary in the policy, not a guideline. Partner and funder requirements on AI use apply in addition to this policy and must be checked before beginning any AI-assisted work on funded projects.

3
In draft policy

Everything AI produces gets checked

Accuracy is the survey's joint top concern. AI generates plausible-sounding content that can be factually wrong. AI tools are also designed to be agreeable, which means they may confirm whatever the user suggests rather than correct it. Every AI contribution to a deliverable is reviewed by the person responsible for it.

4
Survey clarified

Disclosure is proportionate, not blanket

The amount of disclosure required depends on a number of factors: the extent to which AI shaped the content, whether the output is going externally or to a colleague, the nature of the audience, and the sensitivity of the subject. Disclosure is not only about external deliverables: sending AI-generated work to a colleague also warrants a note. The policy sets clear criteria for each context.

5
Survey clarified

The policy enables, it does not restrict

70% of colleagues asked for examples of good practice. That is this policy's primary aim: giving staff the clarity and confidence to use AI effectively across all roles and contexts, not creating compliance anxiety.

6
Survey clarified

We grow alongside AI, not dependent on it

Over-reliance is the co-equal top concern in the survey. The policy names this risk explicitly. Good judgement is not replaced by a prompt, and professional skill development remains essential alongside AI adoption.