Junior employees can be useful guides to generative AI, but a 2024 working paper suggests they should not be an organization’s only source of training or risk advice. In interviews with 78 junior consultants who had used GPT-4, researchers found that proposed safeguards often missed system-level risks. The practical lesson is to pair hands-on experimentation with technical, domain, security, privacy, and governance expertise—not to exclude junior staff.
Why reverse mentoring seemed like a good fit
When a new workplace tool arrives, junior employees may be the first to try it. They are often close to day-to-day workflows, may have fewer established habits around older systems, and can show colleagues practical ways to get started. Peer demonstrations can also feel less intimidating than a formal training session.
Those strengths make junior staff valuable sources of use cases and feedback. They do not automatically make someone an expert in how an AI system fails, what data it exposes, or which controls an organization needs. Generative AI is more than a new interface: its outputs can vary with prompts and context, sound convincing while being wrong, and change as models or integrations change.
What the Harvard, MIT, and Wharton-linked study examined
The working paper, Don’t Expect Juniors to Teach Senior Professionals to Use Generative AI: Emerging Technology Risks and Novice AI Risk Mitigation Tactics, is Harvard Business School Technology & Operations Management Working Paper 24-074, dated June 3, 2024. Its authors were affiliated with Harvard, MIT, Wharton, Warwick Business School, and Boston Consulting Group. The paper is a working paper distributed for comment and discussion, not a controlled trial of workplace training programs. Read the working paper.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →#1 Best Overall
In July and August 2023, junior consultants in a BCG-affiliated business problem-solving exercise used OpenAI’s GPT-4 to work on a fictional apparel company’s channels and brands. Researchers interviewed 78 of them about possible challenges in working with managers and how those challenges might be addressed. Participants generally had one to two years of experience; the managers they discussed had five or more years. The researchers focused on how these less-experienced users reasoned about AI risks, not on comparing age groups or measuring the success of junior-led training. MIT Sloan’s study summary and its discussion of participant experience describe the study context.
Three risk patterns in junior consultants’ advice
1. Treating tool familiarity as reliability expertise
The interviews suggested that some proposed mitigations reflected an incomplete understanding of AI capabilities and limitations, including accuracy, hallucinations, explainability, and contextual relevance. A user who can get a useful answer in one task may not know when the same tool is unreliable, how behavior changes with different prompts or data, or whether an error stems from a user choice or a systemic limitation.
For example, a team member might demonstrate a prompt that produces a polished market summary. That demonstration alone does not establish whether the summary is accurate, whether the underlying information was appropriate to share, or what evaluation is needed before the workflow is used repeatedly. Familiarity with an AI interface is not the same as expertise in AI reliability.
Rank #2
2. Changing people’s routines instead of the system
Recommendations in the interviews often relied on people reviewing prompts and outputs, validating AI-generated work, or agreeing as a team about when AI could be used. Such practices can help, but they leave much of the burden on individual users. They do not by themselves set access permissions, prevent data leakage, log activity, evaluate model performance, secure system integrations, or define what happens when a model or vendor changes.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
“Have a human check it” is a process instruction, not a complete risk-management system. Human review only works when the reviewer has relevant expertise, sufficient time, access to the basis for the answer, and the authority to reject or escalate it. A rushed reviewer who accepts fluent output without checking it is not a meaningful safeguard.
3. Solving for the project rather than the organization
Because participants were close to individual assignments, their proposed solutions tended to remain local. A project team may agree on its own rules and still leave gaps in company-wide data handling, vendor review, security, legal compliance, or consistency across departments.
Risk controls also depend on how a tool is designed and deployed, not just on how one team uses it. Organization-wide acceptable-use rules, data classification, model evaluation, red-teaming, procurement review, and escalation paths address questions that an individual project cannot settle on its own. The paper’s three categories are summarized in its SSRN record.
What the findings do—and do not—say
The study is evidence about novice risk reasoning in a particular early-adoption consulting context. The participants used GPT-4 in 2023 and reflected on risks in interviews. That design reveals the kinds of safeguards they proposed; it does not directly test whether those safeguards would prevent actual incidents.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11- It does not prove junior employees are bad at AI, that younger workers are less capable, or that senior employees know more about AI.
- It did not run a controlled comparison of junior-led and expert-led training, or show that junior-led training caused failures.
- It does not establish that junior staff cannot demonstrate basic prompting, find use cases, collect feedback, or help colleagues learn practical workflows.
- Its sample was junior consultants in a business task, so the findings should not be assumed to apply unchanged to engineers, clinicians, financial professionals, or other work settings.
“Junior” here is principally about experience and expertise, not age. A young specialist may have deep AI knowledge; an older employee may be a novice user. The authors’ broader discussion of the early-implementation context also cautions against treating the finding as a timeless verdict: a later paper on novice risk work.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A safer way to use junior AI champions
Keep junior employees involved, but give them a defined role inside an expert-designed system. They can surface promising use cases, demonstrate approved workflows, test low-risk tasks, gather user feedback, record recurring failures, and escalate uncertainty. They should not be left to own enterprise policy, security architecture, privacy decisions, legal interpretation, model validation, high-impact automated decisions, or final approval of sensitive deployments.
Effective AI education combines different kinds of expertise:
- Domain experts know what a correct result looks like and where an error would matter.
- Technical experts understand models, integrations, data flows, evaluation, and system limitations.
- Security and privacy specialists address access, confidential information, logging, and threat models.
- Legal and compliance teams interpret contractual and sector-specific obligations.
- Learning professionals design training that changes behavior, not just familiarity with features.
- Frontline users and junior staff show where tools fit, where they fail in practice, and what users find confusing.
Teach everyone the basics, then add role-specific depth
All employees need to know which tools are approved, what information they may enter, how to verify outputs, when human review is required, how to report an incident, and which uses are prohibited or restricted. Training should distinguish brainstorming from authoritative analysis: an AI-generated draft may be a starting point, not evidence that its claims are true.
Recommended Free Tools
People using AI in higher-risk workflows need additional instruction in evaluation methods, data handling, reproducibility, model and vendor limitations, bias and fairness, security, regulatory obligations, documentation, and incident escalation. A single prompt workshop cannot cover those responsibilities.
Give champions boundaries and a route to expert help
AI champions need approved tools, a defined scope, a test environment, a standard way to document failures as well as successes, and access to technical and compliance specialists. Give them explicit limits on what they may approve or promise. To avoid driving experimentation into unsanctioned tools, organizations can offer a faster path for low-risk tests while reserving stronger review for sensitive or consequential uses.
Questions to answer before an AI use case becomes routine
- What decision or output will the system influence? Identify whether it is helping with a draft, informing a judgment, or affecting a consequential decision.
- What happens if the output is wrong? The potential harm should determine the level of review and control.
- What information does the system receive or retain? Confirm what data may be entered and what the organization knows about its handling.
- Who is accountable for checking, approving, and correcting the result? Name the responsible person or team and give them authority to stop or escalate the workflow.
If an organization cannot answer those questions, a successful demonstration by a junior employee is not enough evidence to deploy the workflow broadly.
Judge training by behavior, not attendance
Counting attendees or prompts demonstrated says little about whether staff can use AI responsibly. Assess whether employees can spot unreliable output, recognize when not to use AI, follow sensitive-data rules, verify results at a level suited to the task, and report failures. Also check whether training transfers to real workflows, higher-risk uses receive specialist review, and policies are revisited when tools change.
Free tools Windows power users keep installed
One-click scans. No signup required.
Expert-led training is not automatically safe either. Experienced managers and technical specialists can be overconfident or overlook frontline problems. Test training against realistic tasks, assess whether people catch errors, and use feedback from junior champions and other users to improve it.
Keep experimentation open, but assign risk ownership clearly
Excluding junior staff would discard practical workflow knowledge, early signals about usability, and new use cases. Giving them sole responsibility for deciding how AI should be controlled would mistake proximity to a tool for complete risk expertise. Let junior employees help discover where AI may be useful; assign decisions about system controls and consequential uses to people with the relevant technical, domain, security, legal, and governance expertise.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
