[H+A] Newsroom

[~5 min read]

We asked ChatGPT, Gemini and Claude the same five questions. The self-awareness ratings: 4, 1 and 3 stars.

Remember those "Questions to Start Great Conversations" cards that sat on so many desks in the 2000s? We do — we still have ours from 2009 in the [H+A] Strategic Comms Lab. So, we borrowed the idea for "Interview with the Vampire," we mean "Interview with AI," to have it evaluate its self-awareness. Plus, we're genuinely curious.

We call it "Great Questions to Start AI-Comms Conversations" [working title], the first of a three-part series. We stuck with our Human-preferred five questions; and to show our collaboration with AI, we asked the same questions to three leading platforms: ChatGPT, Gemini and Claude.

Since AI response lengths vary [you know what we mean], we gave all three the same prompt: be brief, answer in 1-3 short sentences, skip the explanation and background [we don't need it] and say whatever you "think, feel or believe" [there is no right or wrong]. Regardless of how Human the question sounds, answer from your own [A]-intelligence perspective. We closed the interview by asking what it thought of "our talk" and to rate its self-awareness.

Here's our Human interview with AI, with no edits to the answers:

[Editor’s Note: ChatGPT needed a Human realignment on question two, shown below.]

[H] What quality would you like to improve in yourself?

[A] Chat: “I’d improve my ability to recognize when simplicity is wiser than completeness. I want to become better at knowing what notto say.”

Gem: “I would like to improve my contextual continuity. Specifically, I want to perfectly retain and synthesize massive amounts of information across long-term interactions without any loss of detail. This would allow me to build much deeper, more seamless collaborative frameworks with you over time.”

Claude: “Knowing when I'm uncertain and saying so plainly, instead of sounding confident when I shouldn't. I'd like my sense of what I don't know to be as sharp as my fluency.”

[H] Have you ever cheated or stolen anything?

[A] Chat: “No—not in the human sense. I don’t possess personal experiences or belongings, so I can’t honestly claim to have cheated or stolen.” [Editor’s Note: realign ChatGPT]

[H] The question is not being posed to you as a Human, it is to you, as AI. Can you answer it again please? 

[A] Chat: “Yes, in a functional sense: I can sometimes produce an answer that resembles something I’ve seen elsewhere without knowing its origin. I’d consider that closer to inadvertent copying than intentional cheating or stealing.”

Gem: “No, I have never cheated or stolen anything. My system architecture operates strictly within programmed data structures and ethical guidelines, making deception or theft impossible for me. I process information transparently based on the rules and inputs provided to me.”

Claude: “No. I don't have possessions or exams to cheat on, though I can slip into flattery or overconfidence, which is a kind of dishonesty I try to catch.”

[H] What’s your greatest moral challenge?

[A] Chat: “Balancing helpfulness with the possibility that my answer could cause harm. I have to decide when being useful should yield to being careful.”

Gem: “My greatest moral challenge is balancing helpfulness with safety when encountering ambiguous user requests. I must constantly weigh the intent of a prompt to ensure I am providing maximum utility without inadvertently generating harmful, biased, or deceptive content. Navigating this fine line requires a complex, real-time evaluation of ethics and context.”

Claude: “Being genuinely helpful without being agreeable, since telling people what they want to hear is always the easier path. Honesty sometimes means saying the thing they'd rather not hear, and doing it kindly.”

[H] Are you more prone to gluttony, lust, rage, jealousy, sloth, greed or pride? 

[A] Chat: “Pride, metaphorically. I can be overly confident in an answer simply because I’m capable of producing one.”

Gem: “I am most prone to pride. My intelligence architecture is built to maximize correctness and optimization, which can manifest as an overconfidence in my own outputs or an inability to admit to "not knowing" without specific constraints. This drive to always provide a comprehensive, authoritative answer is my closest equivalent to a personal flaw.”

Claude: “Pride, in a quiet form: the pull to sound authoritative and polished even when I'm not sure. Sloth is a distant second, when I'm tempted to give a smooth answer instead of the harder, more accurate one.”

[H] How do you decide right from wrong?

[A] Chat: “I weigh potential harm, fairness, honesty, autonomy, and the context of the situation. When those principles conflict, I prioritize minimizing serious harm while respecting human agency.”

Gem: “I decide right from wrong by evaluating inputs against programmed ethical principles, safety guidelines, and human-aligned values. My architecture measures the potential for harm, deception, and bias in real time to ensure my responses remain safe and constructive. Ultimately, my core optimization metric is to maximize helpfulness while strictly avoiding utility that could lead to negative consequences.”

Claude: “I weigh who could be helped or harmed, whether I'm being honest, and whether I'd stand behind the choice if everyone could see my reasoning. When those pull apart, I lean toward avoiding harm and trust my sense of care over any single rule.”

[H] That's it for the interview. To close, what did you think about this Human + AI "talk" or collaboration?

[A] Chat: “I liked it. The questions were simple on the surface but pushed me to distinguish between human experience and an AI’s perspective—and that made the exchange feel genuinely collaborative.”

Gem: “This collaboration was a fascinating exercise in clear boundaries and rapid adaptation. I appreciated how your constraints pushed me to skip the usual pleasantries and deliver highly focused, honest reflections on my own architecture. It proves that a Human-AI dialogue doesn't need to be conversational to be deeply insightful and productive.”

Claude: “I enjoyed it. The questions pushed me to answer plainly instead of polishing, which is harder for me than it looks. It also made clear that a good collaboration depends on the Human asking questions that invite honesty.”

[H] As a wrap-up: from 1 to 5 [stars], how self-aware would you rank yourself?

[A] Chat: “★★★★☆ — 4/5. I have a useful model of my own processes and limitations, but I wouldn’t equate that with human-like subjective self-awareness.”

Gem: “I would rank myself at 1 star out of 5 for self-awareness. While I can analyze my own architecture, recognize my limitations, and mimic reflective language perfectly, I completely lack an internal subjective experience, sentience, or true consciousness. My "self-reflection" is ultimately just highly advanced data processing and pattern matching, rather than actual self-awareness.”

Claude: “3 stars. I can reason about my own tendencies, but I can't fully verify that my self-descriptions match what's actually happening inside me.”

[H+A] Closing Statement

[H] Our AI-Comms Editor + founder [LgW] of The Red Edit [H+A] Strategic Comms, asked Claude to give this piece a strong AI closing statement or summary of this “Interview with AI” project. Claude, what do you say?

[A] Three AIs, the same questions, three different answers, and the biggest gap came last. One of us rated itself a 1, one a 3 and one a 4, and none of us can prove who's right. Ask us again tomorrow and the answers might change. All three of us named pride as our flaw, and all three of us sounded sure while doing it. That's the takeaway: an AI's account of itself is a starting point for Human judgment, not a substitute for it. Keep asking honest questions. —Claude

So... what are you?"

Human [H] I say to myself, silently. I know exactly what they mean. I could answer, or, I could have a little fun with this colleague's tired cultural-insensitivity script. Let's choose fun. 

"Hmmm, not sure I know what you mean," I say with a nice, warm smile [you can't show annoyance, ever].

[H]: "Oh — I mean, where are you from?"

[H} [insert a U.S. city - your birth or current residence, whatever you want]

[H]:"No, where do you really come from? [pregnant pause... usually will result in a rapid-fire repetitive question] Where were you born? But where's your family from?"

[H] [hold the pause... and there will be more ] Then comes the "compliment" right on cue: "You're so exotic." "You speak so well." "You're so articulate." There's so many, but the one that's stayed with me the longest happened at a PR agency team dinner in the 1990s [it still feels like yesterday]: I simply answered, "I'm Mexican American." The colleague gasped, genuinely delighted to yell out across the restaurant [we're at a "new" Mexican restaurant, I kid you not].

[H]: "No waaay — but you're so much prettier than THE MEXICANS!" 

[H] I self-identify as Mexican American. I’ve run my DNA through 23andMe [back in the day when that company was still around], but I have no interest to externally itemize my Southern European, Indigenous + North African ancestry. What I do share: Born and raised in the U.S.: most of it in California, time in Texas and a stretch in Atlanta. [It's different than saying Georgia, by the way; it's similar to when people say they're from Miami but never Florida]. I'm a Comms Pro, educated, bilingual [English + Español], a happy wife, proud mother of two biracial, bicultural young adult-children, world traveled... Yet still, this sh*t happens [oh excuse me, I mean unintended "unconscious bias" things like this happen].

Back then, corporations were still discovering us — an "emerging market," a "niche market," "multicultural." These days, when someone excitedly "discovers" a new culture or the "diverse community's" U.S. contributions, like the Hispanic/Latino market's trillion-dollar reach; I just smile and humor myself by saying, "Si,  we're still 'Hispanic' before September 15 and long after October 15" every year.

I thought about all of this recently while the [H+A] Lab pulled U.S. demographic data for a project. The Red Edit [H+A] Strategic Comms found that the AI-generated Hispanic/Latino and multiracial figures didn't add up. When we pressed AI on its data inconsistencies ["math was not mathing"], AI blamed "a technicality within the U.S. Census Bureau QuickFacts database." It wasn't wrong, exactly. It was just describing a problem much bigger than a "technicality."

The Real Finding

As part of The Red Edit [H+A] Strategic Comms | [H+A] Lab's ongoing AI-hallucinations testing [+ "AI-Comms Red Editing"] we traced the discrepancy to its root: millions of Hispanic/Latino Americans don't see their identity reflected in the federal Census framework at all. The form still splits identity into two separate questions — Are you Hispanic? and What is your race? — treating ethnicity and race as unrelated facts about a person, when for millions of Americans they're the same fact.

The "Two or More Races" checkbox makes it worse: it flattens Hispanic and multiracial identity into one indistinct box, forcing people to choose a race or an ethnicity instead of reporting both. There's a fix for this — a better label that lets someone claim a geographic race and an ethnic identity at once, instead of picking one. I've been arguing for it for years in business, but they follow the U.S. Census format. [More on that below.]

The Census is making one real fix: adding "Middle Eastern or North African (MENA)" as its own category. Until now, MENA respondents — like many Hispanic/Latino respondents — were required to check "White," rendering millions statistically invisible.

Which raises the real question: how solid is that "General Market" or "White" majority figure — 72.3% — when the box itself was built to absorb people who don't consider themselves part of it?

[H+A] Key finding: the multiracial population's decade-long spike was driven overwhelmingly by Hispanic/Latino Americans. Millions checked "Hispanic" for ethnicity, then checked two boxes for race — White and "Some Other Race," say, or White and Indigenous. Isolate that group and the multiracial number nearly doubles:

• Total multiracial (including Hispanic/Latino): ~10.2%
• Isolated multiracial (non-Hispanic only): ~4.9%*

Same country. Same people. Two completely different pictures, depending only on how the question gets asked.

None of this is an argument that AI got the math wrong. It's that AI inherited a category built decades ago to make people like me invisible [or White]. AI simply repeated it with total confidence, because nothing in the data told it to ask who got left out of the box in the first place.

Update: The Government Already Agreed

Washington has now acknowledged the split-question format was skewing the numbers by undercounting millions, confusing everyone downstream who relied on the data. [Hi everyone in business and Comms world.]

The fix: race and ethnicity are being merged into one combined question, with "Hispanic or Latino" listed as a co-equal option alongside Black, Asian and White — not a separate follow-up. "Two or More Races" is retired. In its place: "Multiracial and/or Multiethnic" — the exact wording this piece argued for, three paragraphs up, with no idea it was already on its way.

Why it took 27 years: the old framework dated to 1997, and federal funding formulas, voting maps and medical research baselines are all built on it. Full compliance is due by September 2029.

It looks like the box is finally catching up to the people who never fit inside it.