AI can do a lot of things. Comedy isn’t one of them—yet.

Artificial intelligence can summarize research papers, write code, and draft marketing copy in seconds. But ask it to tell a joke, and you’ll often get a flat dad joke or something painfully safe. According to new reporting on internal research from Google DeepMind, even state-of-the-art systems still struggle to be genuinely funny—and that's more than an aesthetic problem. Humor is a social glue. It builds trust, keeps people engaged, and humanizes customer experiences.

In other words: an unfunny AI isn’t just cringe; it can be bad for business.

What DeepMind’s researchers found

Business Insider reported that Google DeepMind researchers recently examined how well their large language models handle humor. Their verdict: today’s models reliably produce safe, formulaic jokes, but they stumble on timing, cultural nuance, and originality. The systems often default to puns, recycle familiar internet gags, and avoid edgier material that many audiences find funny—but that also carries risk.

This isn’t just a Gemini issue. Independent reviewers and researchers have found similar patterns across leading models. As MIT Technology Review put it, AI can write jokes, but that doesn’t mean it’s actually funny.

Why AI struggles with humor

Humor is a high-wire act that blends knowledge, context, timing, and social judgment. A few reasons even cutting-edge models miss the mark:

• Humor is deeply contextual. The same line can be hilarious in one group and offensive—or just confusing—in another. This requires social and cultural calibration, not just word prediction.
• Incongruity and subtext are hard. Many jokes work by setting expectations and delivering an artful twist. That means reading between the lines and anticipating how a human will interpret the setup. Cognitive research shows humor taps multiple systems at once—language, emotion, and social reasoning—making it especially complex for machines to mimic (Frontiers in Psychology).
• Safety training discourages risky humor. Models are tuned to avoid harm and offense. That’s essential, but it also suppresses the transgression and surprise that many forms of humor rely on (Anthropic: Constitutional AI).
• Originality is rare. Models remix patterns from their training data. Without careful prompting or fine-tuning, they often rehash stock formats or known jokes instead of inventing new ones (GPT‑4 Technical Report).

Why this matters for brands and product teams

Humor isn’t just entertainment—it can be a strategic advantage. Used well, it increases recall, improves satisfaction, and takes the edge off tough conversations. Used poorly, it sounds forced, alienates users, or damages trust.

As AI assistants show up in customer support, education, sales enablement, and internal tools, their “voice” increasingly represents your brand. If the assistant can’t read the room—or keeps firing off awkward dad jokes—users notice. Worse, humor that unintentionally targets sensitive groups can create real risk.

That’s the tension: people say they want friendlier, more human AI, but they also want it to be safe and inclusive. Today’s models struggle to balance both.

Safety vs. funniness: the unavoidable trade‑off

It’s not a bug that AI avoids spicy jokes; it’s by design. Leading labs have layered on safety rules to reduce harmful outputs. Google’s AI principles prioritize inclusive, non-offensive behavior (AI at Google: AI Principles). Anthropic’s “Constitutional AI” uses a written set of values to steer models away from risky content (paper).

But comedy thrives on surprise, boundary‑testing, and subverting expectations. When models are trained to play it safe, they predictably become … predictable. That’s why AI humor so often reads as generic, pun‑heavy, and inoffensive to a fault.

How to get better humor from AI today

You can’t turn a chatbot into John Mulaney. But you can coax more engaging, on‑brand humor with structure, guardrails, and testing.

1) Define the comedic lane

• Pick styles that are lower risk but still lively: light observational humor, self‑deprecation, the “rule of three,” or playful callbacks.
• Explicitly ban off‑limits topics (e.g., identity, health, tragedy). Add examples of what to avoid.

2) Give context about the audience

• Share who the content is for, their level of expertise, and the scenario (customer support, onboarding, newsletter, etc.).
• Tell the model what the audience likely knows—and what they don’t.

3) Use structures models handle well

• Prompt for formats like: one‑liner, setup/punchline, analogy with a twist, or a gentle roast of an inanimate object.
• Ask for options and iterate. Keep the best line; discard the rest.

4) Always include a human in the loop

• Review for tone, inclusivity, and originality. Run the content by someone outside the team.
• A/B test with small user groups. If it doesn’t land, pull it.

Example prompts

How researchers are measuring AI humor

Evaluating humor is hard because funny is subjective. Still, the field is building benchmarks and tests to make progress:

• Human ratings at scale. Recruit diverse raters to score jokes on funniness, novelty, and appropriateness—then compare models and humans (the approach described in the Business Insider report).
• Humor detection and rating tasks. NLP challenges like SemEval‑2020 Task 7 evaluate whether systems can detect and rate humor in text. It’s not the same as writing a killer joke, but it tests basic understanding.
• Cognitive plausibility. Some studies look at whether models can explain why a joke is funny, a proxy for whether they grasp the underlying incongruity and resolution (MIT Technology Review).

What could improve next

None of this means AI will never be funny. It means teams need the right ingredients—and constraints—to make incremental gains.

• Better audience modeling. Personalizing tone based on past interactions and stated preferences could help models calibrate humor more precisely.
• Style‑ and persona‑tuning. Fine‑tuning models on curated, inclusive comedic styles (and excluding risky content) can raise the floor without inviting harm.
• Multimodal humor. Visual jokes, memes, and timing cues in audio/video may help AI land simpler, more universal gags.
• Stronger guardrails. Clearer policies, automatic pre‑checks, and human review workflows reduce the risk of a joke backfiring—especially in customer‑facing contexts.

Bottom line

DeepMind’s internal findings echo what many users feel: today’s AI is helpful, polite—and unfunny. That’s okay for most tasks, but it’s a gap for products and brands that rely on warmth and wit. With tighter prompts, careful guardrails, and a human editor, you can get friendlier copy that lands more often than it groans. Just don’t outsource your stand‑up set to a language model.

FAQs

Can AI ever be truly funny?

It can be amusing and even surprising, especially with well‑designed prompts and constraints. But comedy that relies on deep social context and timing remains hard. Expect “pleasantly witty,” not “headline comic.”

Is safety training the reason AI jokes fall flat?

Safety rules do dampen edgier humor, but they’re not the only factor. Humor also needs originality, cultural knowledge, and audience awareness—areas where models are still evolving.

How should businesses use humor with AI today?

Keep it light and situational, define what’s off‑limits, and include human review. Test with small audiences before rolling out broadly.

Will AI replace comedians or copywriters?

No. AI is great for drafts and variations, but the final polish—and the judgment about what’s appropriate—should stay human.

How do researchers measure “funniness” in models?

Mostly with large‑scale human ratings, plus tasks like humor detection and explanation. Technical reports also flag where models fall short in reasoning and nuance.

Sources

1. Business Insider: DeepMind researchers realize AI is really, really unfunny. That's a problem.
2. MIT Technology Review: AI can write jokes now—but can it be funny?
3. Anthropic: Constitutional AI—Harmlessness from AI Feedback
4. OpenAI: GPT‑4 Technical Report (arXiv)
5. SemEval‑2020 Task 7: Assessing Humor in Edited News Headlines
6. Frontiers in Psychology: Humor processing from a cognitive perspective