Credibility First: Professional AI Standards

Responsible AI Use in Professional Services

Part 1: When AI Gets It Wrong, and Who Pays the Price

The lawyers in question. Steven Schwartz and Peter LoDuca of Levidow, Levidow & Oberman, were not reckless people. Schwartz later said he was unaware that ChatGPT could produce false information presented with total confidence. He had used the tool to speed up research, a completely reasonable instinct. The catastrophic error wasn't in using AI. It was in not understanding what AI actually does: it predicts plausible-sounding text, not verified facts. That distinction, invisible to a first-time user, turned a time-saving tool into a professional liability.

This case became a landmark moment for every professional services firm watching from the sidelines, law firms, consulting practices, accounting firms, PR agencies, HR departments. The question it raised wasn't 'should we use AI?' The question was 'what do we need to understand about AI before we use it with client work?' That question is the entire point of this lesson. Responsible AI use in professional services isn't about fear or restriction. It's about knowing where AI is genuinely helpful, where it creates risk, and how to build habits that protect you, your clients, and your firm.

The Principle: AI Is a Confident First Draft, Not a Final Authority

What the Levidow case reveals is a principle that applies far beyond law firms. AI language models, the technology behind ChatGPT, Claude, Google Gemini, and Microsoft Copilot, are trained to produce fluent, coherent, contextually appropriate text. They are extraordinarily good at sounding authoritative. They are not, by design, fact-checkers or research databases. Think of it this way: if you hired a brilliant new associate who had read almost everything ever published but had no way to verify what they were recalling, you'd be impressed by their fluency but you'd check their work before sending it to a client. That's exactly the relationship you should have with AI-generated content.

This matters differently depending on your role. A marketing manager using Claude to draft social copy has relatively low stakes, a wrong statistic in a LinkedIn post is embarrassing but fixable. A management consultant using ChatGPT Plus to draft a client-facing market analyzis has much higher stakes, a fabricated market size figure presented to a CFO can destroy a relationship and expose the firm to liability. A financial advisor using Microsoft Copilot to summarize regulatory guidance has the highest stakes of all, acting on a hallucinated regulatory requirement could harm clients and trigger compliance investigations. Same tool, same underlying behavior, vastly different consequences.

The term 'hallucination' has become the standard shorthand for when AI generates false information with confidence. It's a useful word, but it can be misleading, it implies something rare and dramatic, like a one-off glitch. In reality, AI hallucinations exist on a spectrum. At one end, you have outright fabrications like the invented legal cases. At the other end, you have subtle distortions: a statistic that's slightly out of date, a company description that was accurate two years ago, a regulatory summary that omits a key exception. The subtle ones are often more dangerous because they pass casual review. Building a habit of verification isn't paranoia, it's professional hygiene.

A Different Industry, The Same Risk: PR and the Fabricated Quote

Move from law to public relations, and the responsible AI challenge looks different on the surface but identical underneath. In 2023, Sports Illustrated was found to have published articles under fake author names, with AI-generated bylines complete with fabricated biographies and headshots generated by image AI tools. The articles themselves contained content errors and generic filler that editors would normally catch. When the story broke, reported by Futurism, the backlash was swift. Advertisers asked questions. Subscribers canceled. The editorial team faced internal investigations. Sports Illustrated's publisher, The Arena Group, eventually terminated its licensing agreement. The magazine's century-long brand took damage that a correction notice couldn't undo.

For PR and communications professionals, this case illustrates a specific failure mode: using AI to replace human judgment rather than support it. A communications manager using Grammarly AI to polish a press release, or using ChatGPT Plus to draft three versions of an executive quote for approval, is using AI responsibly, the human remains in the loop, making decisions. A team that publishes AI output without review, attribution review, or fact-checking has removed the professional from professional services. The tool didn't fail. The workflow did. And in a field where credibility is the entire product, workflow failures are brand failures.

Comparing Risk Levels Across Professional Services Roles

Click to enlarge

A Third Example: The HR Manager and the Biased Job Description

The risks of unreviewed AI output aren't always about factual accuracy. Sometimes the problem is bias baked into the model's training data. In 2018, before the current generation of AI tools. Amazon scrapped an internal AI recruiting tool after discovering it systematically downgraded résumés from women. The model had been trained on a decade of Amazon's historical hiring data, which reflected the male-dominated patterns of the tech industry. By 2024, HR teams were encountering similar issues with AI-generated job descriptions: language models trained on historical job postings tended to reproduce gendered phrasing, age-coded language, and requirements that excluded protected groups without any conscious intent from the HR professional using the tool.

An HR manager using ChatGPT to draft a job description for a 'rockstar developer who thrives in a fast-paced, high-energy environment' might not realize that research consistently shows this kind of language skews male applicant pools. Or that requiring '5+ years of experience' for a role that could be done by a two-year professional may constitute age-related screening that creates legal exposure in some jurisdictions. The AI didn't intend any of this, it generated language that matched patterns in its training data. The HR professional who reviews that draft with an eye for bias, legal compliance, and inclusivity is doing their job. The one who copies and pastes it directly into the applicant tracking system is not.

Practical Application: Building Verification Into Your Workflow

Responsible AI use isn't a philosophy, it's a set of habits. The most effective consulting teams and professional services firms that have integrated AI tools successfully share one common practice: they treat AI output as a first draft that needs a specific type of review, not a general proofread. That means identifying, before you even open ChatGPT or Claude, what category of content you're asking for. Structural content, outlines, frameworks, draft agendas, email templates, carries low verification burden. Factual content, statistics, regulatory references, company details, case studies, financial figures, carries high verification burden. Creative content, tone, phrasing, messaging options, sits in the middle, requiring human judgment about appropriateness rather than fact-checking.

Microsoft Copilot, embedded in Word and PowerPoint, makes this distinction particularly important because the seamless interface can create a false sense of security. When Copilot summarizes a document for you, it's working from the actual document, lower hallucination risk. When you ask it a general knowledge question or ask it to 'add some industry context,' it's drawing on its training data, higher hallucination risk. Google Gemini in Google Workspace behaves similarly: document-grounded tasks are more reliable than open-ended knowledge generation. Knowing which mode your AI is operating in is a practical skill, not a technical one. It just requires asking: is this tool working from something I gave it, or generating from memory?

The firms getting this right are also building team-level habits, not just individual ones. That looks like a simple internal policy: AI-generated content must be labeled as such in internal drafts until it's been reviewed. Client-facing materials require a named human reviewer who signs off before delivery. Any statistic or regulatory claim sourced via AI must have a backup primary source link in the document notes. These aren't bureaucratic hurdles, they're the same quality controls that good firms already apply to junior associate work. AI is, in effect, a very fast, very fluent, very junior contributor. Treat its output accordingly.

Key Lessons So Far

1. AI tools generate fluent, confident text, not verified facts. Fluency is not accuracy.
2. The Levidow legal case shows that even experienced professionals can mistake AI output for reliable research. The error isn't in using AI, it's in skipping verification.
3. Risk scales with stakes. The same AI tool carries different consequences for a marketing manager versus a financial advisor versus a litigating attorney.
4. Hallucinations exist on a spectrum. Subtle errors, a slightly wrong statistic, an outdated regulatory detail, are often more dangerous than obvious fabrications because they pass casual review.
5. Bias is a form of AI error too. Job descriptions, hiring criteria, and policy drafts generated by AI can reflect historical biases without any intent from the user.
6. AI operating on documents you provide is more reliable than AI generating from general training data. Knowing which mode your tool is in is a practical skill.
7. Verification habits need to match the content type: structural content requires less scrutiny than factual or compliance-related content.
8. Responsible AI use is a team practice, not just an individual one. Firms that get this right build shared policies around labeling, review, and sign-off, just as they do with junior associate work.

When Confidentiality Meets the Chatbot

In early 2023, Samsung engineers made a decision that seemed perfectly reasonable at the time. They were debugging proprietary semiconductor code and needed help fast. ChatGPT was right there. They pasted the code in, got useful answers, and moved on. Within weeks, Samsung's internal security team discovered what had happened: confidential source code, internal meeting notes, and hardware specs had been entered into a public AI system. Samsung responded by banning employee use of generative AI tools on company devices. The engineers weren't being reckless, they were being efficient. That's exactly the problem.

The Samsung incident became a cautionary tale across professional services almost overnight. Law firms, consulting houses, and financial advisory firms started asking a question they hadn't thought to ask before: when an employee pastes client data into an AI tool, where does that data actually go? The answer depends entirely on which tool, which plan, and which settings, and most professionals have no idea. The default behavior of many consumer AI tools is to use your conversations to improve the model. That means client information you type could, in theory, become training data.

The principle hiding inside the Samsung story isn't about technology. It's about professional duty. Consultants, lawyers, HR advisors, and accountants operate under confidentiality obligations that predate AI by decades. Those obligations don't pause because a tool is convenient. What Samsung's engineers lacked wasn't skill, it was a clear mental model for which information can leave the building and which cannot. Building that mental model is one of the most immediately practical things you can do as a professional who uses AI.

The Three-Zone Framework for Information

A simple mental model used by several professional services firms divides workplace information into three zones. Zone One is public or generic information, industry trends, general frameworks, publicly available data, your own thinking about a problem. This can go into any AI tool without concern. Zone Two is internal but non-sensitive, draft agendas, general process descriptions, anonymized examples, your own professional development work. This can go into enterprise-grade AI tools with appropriate controls. Zone Three is confidential, named client data, financial figures tied to a specific organization, personnel records, legal matters, proprietary strategy. This stays off AI systems unless your firm has a specifically approved, secure deployment.

Most professionals find that the vast majority of their AI use falls comfortably in Zone One and Zone Two. The confusion, and the risk, comes from Zone Three material sneaking into prompts without the person fully registering what they've included. A consultant drafting a slide deck might paste in a client's revenue figures to help frame a narrative. An HR manager might describe a specific employee situation in detail to get advice on handling it. A lawyer might summarize a client's legal exposure to get help with a brief. Each of these feels like using AI to do your job. Each of them is also a potential confidentiality breach.

The fix isn't to stop using AI. The fix is to develop the habit of anonymizing before you paste. Replace client names with 'Client A.' Replace specific revenue figures with illustrative ranges. Describe a personnel situation in structural terms rather than naming the individual. You lose almost nothing in terms of AI output quality, the model doesn't need the real names to help you think through the problem, and you maintain the confidentiality standard your role requires. This is a discipline that takes about two weeks of deliberate practice to become second nature.

Click to enlarge

A Law Firm's Lesson in Verification

In June 2023, a New York attorney named Steven Schwartz submitted a legal brief that cited multiple court cases in support of his client's position. The problem: several of those cases didn't exist. They had been generated by ChatGPT, which fabricated plausible-sounding but entirely fictional case citations, complete with realiztic case names, docket numbers, and judicial opinions. Schwartz had used ChatGPT to research precedents and hadn't independently verified what the tool produced. The judge was not sympathetic. Schwartz and his firm faced sanctions. The case became front-page news.

This phenomenon. AI generating confident, detailed, completely wrong information, has a name: hallucination. It happens because large language models don't look things up the way a search engine does. They predict what text should come next based on patterns in their training data. When asked for a specific case citation or a precise statistic, a model will sometimes generate something that fits the pattern of what a real answer would look like, even when it has no actual knowledge of the specific fact. The model doesn't know it's wrong. It produces the invented citation with exactly the same confident tone it uses when it's completely accurate.

The Hallucination Problem Across Different Roles

The risk of AI hallucination isn't unique to law. A financial consultant who asks an AI to summarize a company's recent earnings performance might get a confident summary that mixes up two different quarters or invents a metric. A marketing consultant who asks for statistics about consumer behavior in a specific sector might get a plausibly formatted research citation that doesn't actually exist. An HR advisor who asks about employment law in a specific jurisdiction might get a confident description of regulations that haven't been in force for three years, or that apply in a different state entirely.

The pattern is consistent across roles: AI tools are most reliable when they're helping you think, structure, and draft, and least reliable when they're being asked to recall specific facts. A useful rule of thumb: if you could verify the claim in under two minutes with a Google search or a database lookup, do it. If you couldn't easily verify it, that's a sign you should be even more careful about relying on it. The professional risk isn't the AI being wrong, it's you presenting the AI's wrong answer as your professional judgment.

An HR Director's Bias Blind Spot

A large UK retailer piloted an AI-assisted screening tool for store manager candidates in 2022. The HR team fed the system historical hiring data and asked it to identify candidates with profiles similar to their highest-performing managers. The tool worked efficiently, and then someone noticed the pattern. The system was systematically ranking candidates from certain postcodes lower, and candidates with non-Western names lower. It wasn't doing this randomly. It was doing it because the historical data it learned from reflected the same biases that had shaped past hiring decisions. The tool had industrialized a bias that humans had been applying inconsistently.

This is a different kind of AI risk from hallucination, and it's particularly relevant for HR professionals, managers making staffing decisions, and consultants who help clients design talent processes. AI tools learn from historical data. If historical decisions reflected systemic biases, conscious or not, the AI will reproduce and potentially amplify those patterns. The risk isn't that the AI is malicious. The risk is that it makes biased decisions at scale, consistently, and with an air of algorithmic objectivity that makes the bias harder to see and challenge. Human reviewers who might have caught an individual biased decision get bypassed entirely.

Turning Principles Into Daily Habits

Responsible AI use in professional services isn't a compliance exercise you complete once. It's a set of habits that become part of how you work. The professionals who handle this best aren't the ones who memorize the most rules, they're the ones who've built a small number of instinctive checks into their workflow. Before they paste anything into an AI tool, they've already asked: does this contain client-identifiable information? Before they include any AI-generated figure or citation in a deliverable, they've already asked: have I verified this independently? Before they use AI output to make a decision about a person, they've already asked: am I the accountable decision-maker here?

These checks take seconds when they're habits. They take seconds to skip when you're rushing. The professional risk of skipping them isn't abstract, it's the Samsung situation, the Schwartz situation, the UK retailer situation. Each of those cases involved competent professionals making a fast decision that seemed reasonable in the moment. The discipline isn't about being slow or overly cautious. It's about being as rigorous about AI-assisted work as you already are about other professional outputs that carry your name.

The good news is that none of this requires you to become a technical expert. You don't need to understand how large language models work at a mathematical level to use them responsibly. You need to understand what they're good at, where they fail, what your professional obligations require, and which tools your firm has approved for which types of work. That knowledge is entirely accessible to non-technical professionals, and increasingly, it's the knowledge that distinguishes professionals who use AI effectively from those who either avoid it entirely or use it carelessly.

Key Lessons From This Section

1. The Samsung incident shows that confidentiality risk in AI use is almost always a habit problem, not a knowledge problem, professionals paste sensitive information into tools without stopping to categorize it first.
2. The three-zone framework (public/generic, internal/non-sensitive, confidential) gives you a practical mental model for deciding what can and cannot go into an AI tool.
3. Anonymizing client data before using AI assistance preserves nearly all of the tool's usefulness while maintaining your professional confidentiality obligations.
4. AI hallucination, confident, detailed, completely wrong output, is a structural feature of how language models work, not a bug that will be fixed. Your verification habits are the only reliable defense.
5. The Schwartz case is the clearest professional consequence available: presenting AI-generated factual claims without verification can result in sanctions, reputational damage, and client harm.
6. AI tools are highly reliable for reasoning, structuring, and drafting tasks, and significantly less reliable for recalling specific facts, statistics, citations, and recent events.
7. AI bias in people-related decisions is a systemic risk, not an edge case. Tools trained on historical data will reproduce historical patterns, including discriminatory ones, at scale.
8. Responsible AI use is a set of daily habits, three instinctive checks applied consistently, not a compliance framework you revisit annually.

When AI Gets It Wrong: Managing Risk, Bias, and Accountability in Client Work

In 2023, two lawyers at the New York firm Levidow, Levidow & Oberman submitted a legal brief citing six court cases that did not exist. ChatGPT had fabricated them, complete with plausible-sounding case names, judges, and ruling dates. The lawyers had not verified a single citation. The judge fined them $5,000 each and ordered them to notify the judges supposedly named in the fake rulings. The story made international headlines, but the underlying mistake was not exotic. It was the same error any consultant, marketer, or HR professional could make: trusting AI output without checking it against reality.

What made this case instructive was not the lawyers' incompetence, they were experienced practitioners. It was the seductive confidence of AI-generated text. ChatGPT does not hedge or stammer. It presents fabricated citations with the same tone and formatting as real ones. That fluency is precisely what makes verification non-negotiable. The principle extracted from this story is simple: AI is a drafting tool, not a source of truth. Every factual claim it produces, statistics, case references, client data interpretations, must be independently confirmed before it reaches a client.

For consultants, the stakes extend beyond embarrassment. Client relationships depend on trust. A fabricated market statistic in a board-level presentation, an inaccurate regulatory reference in a compliance report, or a misattributed quote in a due diligence document can cost a firm its reputation far faster than any missed deadline. The good news: the fix is not complex. It requires a habit, not a technology upgrade, a consistent practice of treating AI output as a first draft that needs a human expert's eye before it becomes client-facing work.

Bias in the Brief: When AI Reflects the Past, Not the Possible

Amazon built an AI hiring tool in 2014 to screen resumes faster. By 2018, they had quietly scrapped it. The system had been trained on a decade of historical hiring data, data that reflected a predominantly male engineering workforce. It learned to penalize resumes that included the word 'women's' (as in 'women's chess club') and downgraded graduates of all-women's colleges. Amazon's own engineers could not fix the bias, because the bias was baked into the training data itself. The tool was not malfunctioning. It was doing exactly what it was designed to do: replicate past patterns.

For consultants advising clients on AI adoption, or using AI tools in their own HR, talent, or organizational work, this story carries a direct warning. AI tools trained on historical data will encode historical inequities. A workforce planning AI trained on past promotion data may systematically undervalue profiles that look different from those previously promoted. A sentiment analyzis tool trained on Western English-language data may misread responses from global teams. The bias is invisible unless you look for it.

The practical implication is not to avoid AI in sensitive work, it is to ask the right questions before deploying it. Who was this tool trained on? What data did it learn from? Does that data reflect the population I am now applying it to? These are not technical questions. They are strategic ones, and they belong in every consulting engagement where AI touches people decisions.

Confidentiality Failures Happen Before You Notice

In early 2023, Samsung engineers used ChatGPT to help debug proprietary source code. On three separate occasions, employees pasted confidential data, including internal meeting notes and unreleased semiconductor designs, directly into the chat interface. Samsung had not yet established any AI usage policy. Within weeks, the company banned ChatGPT on corporate devices entirely. The data had already been submitted to OpenAI's servers, potentially used to train future models. Samsung could not retrieve it.

For consulting professionals, the Samsung case is not a cautionary tale about technology. It is a cautionary tale about habits formed before policies exist. Consultants routinely handle sensitive client information: financial projections, M&A targets, employee performance data, unreleased product plans. The instinct to paste a messy document into ChatGPT for a quick summary is natural, and potentially a serious breach of client confidentiality agreements. The question to ask before every AI interaction is: if this text appeared in a news article tomorrow, would my client be harmed?

The Teacher Who Built the Right Habit

A high school principal in Ontario began using Claude to draft parent communications, staff memos, and board reports in 2023. Early on, she developed a two-step rule she called 'generate, then govern.' She would let the AI produce a full draft, then spend five minutes asking herself three questions: Is every fact here verifiable? Does this reflect the full diversity of our school community? Would I be comfortable if a parent saw exactly how this was made? The habit took less time than her previous editing process and caught two errors in the first month, a misquoted policy reference and a demographic assumption about family structures that did not reflect her school's reality.

What this principal built was not a compliance framework. It was a professional reflex. The same reflex applies directly to consulting work. Before any AI-assisted deliverable leaves your hands, run the 'generate, then govern' check: verify the facts, examine the assumptions, and confirm you would stand behind the process if the client asked how it was made. Most of the time, the answer will be yes. Occasionally, it will save you from a serious mistake.

Click to enlarge

Building Responsible AI Habits Into Your Consulting Practice

Responsible AI use in professional services is not about fear. It is about professional standards applied to a new tool. Doctors verify drug interactions even when a pharmacist flags them. Lawyers cite-check even when a paralegal has already reviewed. Consultants model-check even when a senior analyzt ran the numbers. Adding AI to the workflow does not lower the verification standard, it raises the need for it, because AI errors are more confident-sounding and harder to spot than human ones.

The firms building durable AI practices right now are not the ones using AI most aggressively. They are the ones with clear internal guidelines: what data can go into which tools, which deliverables require human expert review, and how AI assistance is disclosed to clients when relevant. Some firms, including Deloitte and PwC, have deployed enterprise-licensed versions of AI tools specifically so that client data stays within controlled environments rather than public model training pipelines.

You do not need to wait for your firm to build a policy before you build personal habits. Start with three commitments: never paste identifiable client data into a consumer AI tool, always verify specific facts before they reach a client, and treat AI output as a starting point that requires your professional judgment before it becomes advice. Those three commitments separate professionals who use AI well from those who eventually get burned by it.

Key Lessons from This Section

1. AI hallucination is a structural feature of large language models, not a bug to be patched, every specific fact in AI output requires independent verification before it reaches a client.
2. Bias in AI tools often reflects historical data patterns; consultants advising on people decisions must ask vendors directly about training data demographics and disparate impact testing.
3. Pasting confidential client data into consumer AI tools is a confidentiality risk, enterprise-licensed tools with controlled data environments exist precisely to address this.
4. The Samsung and Levidow cases show that AI risk incidents happen before policies exist; personal habits must precede firm-wide frameworks.
5. The 'generate, then govern' reflex, verify facts, examine assumptions, confirm you stand behind the process, is the core professional habit for responsible AI use.
6. Responsible AI use is a competitive differentiator in consulting: clients are increasingly asking how firms handle AI and data, and clear protocols build trust.

Takeaways

• Never send AI-generated facts, statistics, or references to a client without tracing them to a primary source first.
• Treat AI output as a talented junior analyzt's first draft, useful, fast, but unverified until you check it.
• Consumer AI tools are not safe containers for real client data; use anonymized placeholders in prompts.
• Bias in AI hiring and people tools is real, documented, and your professional responsibility to flag when advising clients.
• A five-check pre-send habit takes two minutes and protects your professional reputation.
• Personal protocols do not require a firm policy to exist, build yours now.