How a Museum Marketing Team Used AI to Bring People Closer to Art
Harvard Business Review · podcast · 2026-07-21 source ↗ Reader 🖥️
AI BacklashAuthenticity Premium (Human Signal in an AI World)Decision-Making PracticesEnterprise AI Adoption & ReadinessTrust
Drawing on an HBS case study, HBS professor Julian De Freitas and Harvard Art Museums marketing director John Connolly use the Museums' 2025 AI-avatar campaign to argue that adopting generative AI in a public-facing, brand-sensitive context is not a generic technology-rollout decision but a domain-specific one: AI's threat to perceived authenticity is sharpest in domains built on one-of-a-kind originals, high trust, or expected human touch (art, tax preparation, healthcare, education), so organizations in those domains must weigh interpretive authority, stakeholder trust, and brand license before novelty or attention. The case shows this weighing playing out procedurally, not just conceptually: Connolly deliberately built face-to-face (not email) consensus with skeptical curators, conservators and art historians, framed the initiative obliquely ('45 degrees') rather than announcing 'let's do AI', and treated radical transparency about AI's use as the load-bearing safeguard against the exact authenticity risk the technology raises — a governance pattern the speakers suggest generalizes to any brand deciding whether it has 'license' to use a given AI application versus merely appearing to jump on the bandwagon.
practical · mixed — A Cold Call podcast interview (HBR) built around an HBS case, 'Harvard Art Museums: When Art Meets Artificial Intelligence'. Primarily practical/argumentative: HBS professor Julian De Freitas and Harvard Art Museums marketing director John Connolly extract generalizable lessons about how organizations should weigh and execute risky generative-AI adoption decisions in brand-sensitive contexts, illustrated through narrative recounting of the museum's internal stakeholder debate and campaign rollout.
Structure
| Part | Role |
|---|---|
| Opening framing: museums as guardians of history/tradition facing pressure to attract younger, more distracted audiences | Sets up the innovation-versus-stewardship tension the episode explores |
| De Freitas explains why he wrote the case: AI is a category apart in consumer perception, and art museums are an extreme test case because the product (a one-of-a-kind original) can't be duplicated | Establishes the theoretical frame (domain-specific AI adoption barriers) that the rest of the conversation applies |
| Connolly's institutional context: a 2025 shift to free admission removed a membership incentive, forcing new engagement/revenue strategies in an extremely crowded competitive market | Context establishing why the museum felt pressure to take a marketing risk at all |
| The proposal itself: an AI-animated video avatar of a canonical artwork plus a two-way AI chatbot version | Names the concrete technology decision under debate |
| Stakeholder pushback mapped onto roles — curators, conservators, art historians, museum leadership — and generalized by De Freitas into a theory of when AI provokes resistance | Core evidence for the domain-specific-barriers thesis |
| The Barbie-box campaign recounted as institutional precedent for boundary-pushing marketing | Narrative evidence that curator-approved risk-taking had already succeeded once, lowering the bar for the AI proposal |
| Connolly's internal buy-in process: face-to-face conversation (not email), oblique '45-degree' framing, and a hard transparency commitment (never covert AI, never in donor communications) | Core evidence for the governance-process thesis |
| De Freitas's generalizable framework: domain-specific barriers plus execution-specific skepticism (performative AI use vs. genuine, otherwise-unattainable value) | Synthesizes the case into transferable lessons for other brands |
| Outcome: mixed but net-positive reception, negative feedback tracking general AI backlash rather than museum-specific harm; campaign data being analyzed for publication | Reports the result without a rigorous counterfactual |
| Closing reflection on whether these tensions are transitional or durable, with implications for brand positioning and policy | Leaves the central open question unresolved deliberately |
Key terms
- brand license
Whether an audience judges a brand as credibly entitled to use a given technology or make a given claim, versus perceiving it as opportunistically jumping on a bandwagon. - domain-specific AI adoption barriers
De Freitas's framework that resistance to AI is not uniform across use cases but concentrated in domains involving one-of-a-kind originals, high-trust requirements (e.g. tax preparation), or expected human touch (healthcare, education) — success in one domain does not predict acceptance in another. - 45-degree framing
Connolly's tactic of introducing a controversial AI initiative by leading with its purpose ('bring people closer to art') rather than announcing 'we're doing AI', to avoid triggering reflexive resistance before the goal is understood. - performative AI use
Using AI in a way perceived as superficial, cost-driven, or investor-pleasing rather than adding real, otherwise-unattainable value — the pattern De Freitas says audiences are increasingly skeptical of.
Propositions
- explicitAI is a fundamentally different technology from a consumer-perception standpoint, which is why it creates unique sources of value, unique risks, and unique barriers to adoption compared to other new technologies an organization might adopt.conf: high
- explicitAI adoption barriers are domain-specific, not uniform: they are especially salient where the product is a one-of-a-kind original (art), where trust is paramount and errors are unacceptable (e.g. tax preparation), and where human touch is expected (healthcare, education) — success with AI in one domain does not imply it will be accepted in another.conf: highextendsAdam Mosseri: AI is a tailwind for authenticity · p20 — Mosseri's general claim that abundant synthetic content drives demand for authenticity/humans is extended here with a concrete mechanism: De Freitas names the specific domain categories (one-of-a-kind originals, high-trust, human-touch) where that authenticity premium is sharpest.
- explicitAudiences are increasingly skeptical of AI adoption that looks performative — cost-cutting or investor-pleasing dressed up as innovation — and instead judge whether the organization is doing something with AI it genuinely could not have done otherwise, delivering real customer or educational value.conf: high
- explicitWhether an audience accepts a brand's use of a given AI application depends on whether the brand is perceived to have 'license' to make that technology claim given its history and associations, versus being seen as opportunistically jumping on the AI bandwagon — so the authenticity discussion has to be had freshly for each brand and audience rather than assumed from other cases.conf: high
- explicitRadical transparency about AI use — never concealing that content is AI-generated, and categorically refusing to use AI in trust-sensitive channels like donor communications — was treated as the key safeguard that let curators sign off, because the specific danger the team feared was audiences discovering covert AI use and being pushed further away from an institution already working to close a historic public-trust gap.conf: highsupportsAdam Mosseri: AI is a tailwind for authenticity · p21 — Mosseri argues content should be judged by provenance transparency rather than concealment; Connolly's categorical commitment to overt AI disclosure (and refusal to use AI covertly in donor communications) is a real-world institutional instance of exactly that provenance-first stance.
- explicitInternal buy-in for a risky AI initiative was deliberately built through face-to-face conversation rather than email (avoiding a CC'd paper trail), and by leading with the initiative's purpose rather than announcing 'we're using AI' — an oblique framing intended to let stakeholders engage with the goal before reacting to the technology.conf: high
- explicitDifferent internal stakeholder groups objected to the AI avatar for distinct, role-specific reasons: curators worried about misrepresenting artist intent, conservators about altering the artwork's integrity, art historians about de-skilling the public's ability to analyze art independently, and museum leadership about the avatar being perceived as an official museum spokesperson that could hallucinate facts or lose control of the institution's narrative.conf: high
- implicitAn existing track record of surviving smaller, curator-approved boundary-pushing marketing (e.g. the Barbie-box social campaign) functioned as evidence within the institution that provocative marketing experiments could generate engagement without lasting reputational harm, making it easier to propose a larger AI-based risk.conf: medium
- explicitThe campaign's actual outcome was mixed but net favorable: strong engagement and appreciation for the technology, some negative reaction, but the negative feedback tracked general public anxiety/backlash about AI rather than a museum- or Harvard-specific reputational failure.conf: high
- explicitIt remains unresolved whether AI-adoption resistance in domains like art is a transitional friction that will dissolve as the technology becomes familiar, or a durable feature of domains where people will keep insisting on human presence — a distinction with implications for how brands position themselves and whether policy should protect a 'right' to human touch in some domains.conf: high
Arguments
AI adoption should be treated as domain- and execution-contingent rather than as a single go/no-go call generalizable across an organization, and it remains an open question whether resistance in authenticity-sensitive domains is transitional or durable. strength: medium
Explicit premises:
- AI is a fundamentally different technology from a consumer-perception standpoint
- AI adoption barriers concentrate in domains with one-of-a-kind originals, high trust requirements, or expected human touch
- Audiences penalize AI use that looks performative rather than genuinely value-adding
- Acceptance of a brand's AI use depends on perceived 'license', which must be assessed fresh per brand rather than inferred from other cases' success
Implicit premises:
- Organizations lack an a priori way to know which domain category they fall into before running the experiment
- Consumer perception of AI, not the technology's technical capability, is the binding constraint on adoption
Inferences:
- A brand's AI-adoption playbook cannot simply be copied from another brand's successful case, even within the same industry
- Institutions in trust- or authenticity-sensitive domains should expect to re-litigate the adoption decision for every new AI use case rather than treating a first success as blanket permission
Overt, categorical transparency about AI use, combined with a deliberately built face-to-face internal-consensus process, is what converted a high-authenticity-risk AI initiative into one the institution's stakeholders could accept and that ultimately produced a net-positive outcome. strength: medium
Explicit premises:
- Transparency about AI use (never concealing it, never using it in donor communications) was explicitly positioned to curators as the safeguard against authenticity risk
- Buy-in was built face-to-face rather than by email, avoiding a documented trail and allowing nuance in framing
- Stakeholder objections were role-specific and were addressed individually rather than with a single blanket justification
- The eventual outcome was net favorable, with negative feedback attributable to general AI sentiment rather than the museum's specific execution
Implicit premises:
- Stakeholders who might reflexively reject 'AI' in the abstract can be brought to accept a specific, well-governed instantiation of it
- The causal link between the transparency commitment and the outcome is real rather than coincidental with other factors (novelty, execution quality, curator sign-off itself)
Inferences:
- Institutions considering similarly risky AI use cases should build explicit, non-negotiable transparency commitments into the proposal before seeking stakeholder buy-in
- The process of building consensus (channel, framing order, stakeholder-specific engagement) may matter as much to outcomes as the technology decision itself
Narratives
- Rather than circulating the AI-avatar proposal by email, Connolly met with curators and other stakeholders face-to-face, deliberately avoiding a CC'd paper trail. He introduced the idea 'at a 45-degree angle' — leading with the goal of bringing people closer to art rather than announcing an AI initiative outright — and foregrounded total transparency about the technology's use, including a hard rule never to use AI in donor communications. The curators, initially wary, ultimately signed off after these conversations, and one curator was reportedly startled and impressed when shown the avatar demo on Connolly's phone.
→ Illustrates the governance/consensus-building process behind p5 and p6 — the mechanics of how an institution actually converts stakeholder resistance into buy-in for a risky AI initiative. - Before the AI avatar campaign, the museum's social media strategist ran a comedic Instagram post positioning an ancient artifact inside a mock Barbie box, developed in collaboration with curators. It drew some 'this is a travesty' backlash in comments alongside strong positive engagement, and Connolly cited it as evidence that provocative, boundary-pushing marketing, done with curator sign-off, could succeed without lasting damage — precedent that made the larger AI risk easier to propose.
→ Supports p8 — an internal precedent of successful risk-taking lowered the perceived risk of proposing the larger, more sensitive AI initiative.
Questions addressed
- ⭕ How should an organization decide whether it has 'license' to use a specific, risky AI application given its brand and history?
De Freitas names the judgment ('brand license' vs. bandwagon-jumping) and says it must be assessed fresh per brand, but offers no decision procedure for making that assessment — it is described, not operationalized. - ✅ Which domains and product types make AI adoption especially fraught from a consumer-authenticity standpoint, and why?
One-of-a-kind originals, high-trust domains (e.g. tax preparation), and domains expecting human touch (healthcare, education) are named as especially resistant, because AI use there raises authenticity, trust, or human-presence expectations that other domains don't. - ✅ What internal process (stakeholder engagement, framing, transparency commitments) converts AI-adoption resistance into buy-in inside a risk-averse institution?
Face-to-face (not email) engagement, oblique goal-first framing, and a categorical transparency commitment are described as the process that won curator sign-off. - ⭕ Is resistance to AI in trust- or authenticity-sensitive domains a temporary adjustment period or a durable preference for human involvement?
De Freitas explicitly declines to resolve this, framing it as an open question with policy implications (whether a 'right' to human touch should be guaranteed in some domains) rather than offering an answer.
Gaps
- incompleteNo criteria are given for how the museum weighed the avatar campaign's benefits (attention, engagement, membership) against its risks (donor relations, reputational, Harvard-brand association) — the episode describes the process (face-to-face conversation, transparency commitment, oblique framing) but never surfaces an explicit decision framework, cost/benefit threshold, or who held final go/no-go authority once curators signed off.
Read next: Structured decision-making frameworks for weighing novel-technology risk against brand value (e.g. curated options, pressure-testing, calibration) applied to a real go/no-go case, to see whether Connolly's process maps onto or diverges from formal decision-science practice. - incompleteDe Freitas frames domain-specific AI resistance (originals, trust-heavy domains, human-touch domains) as an emerging finding, but the episode gives no account of how an organization would determine in advance, before running the experiment, whether its own domain falls into a resistant category — the taxonomy is diagnosed only in hindsight from cases that have already been run.
Read next: Research or frameworks that predict, ex ante, which product/brand domains will trigger authenticity-based AI resistance, rather than diagnosing it post hoc. - uninformedThe claim that transparency (never concealing AI use) mitigated authenticity risk is asserted from a single case with a self-reported, undetailed 'mixed but positive' outcome; no comparison is offered against a counterfactual (e.g. a covert or partially-disclosed AI campaign) that would let a listener assess how much of the outcome is attributable to transparency specifically versus other factors (curator buy-in, novelty, campaign execution).
Read next: Controlled studies comparing disclosed vs. undisclosed AI-generated brand content on trust/authenticity outcomes, to isolate the transparency variable from confounds like novelty and creative quality.
Related documents
- Adam Mosseri: AI is a tailwind for authenticity
- AI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy
- Brain Food: Building Trust in Yourself Every Day
- Stop being skeptical about AI for development with Charity Majors
- Generative AI in the Real World: Aishwarya Naresh Reganti on Making AI Work in Production
- Why "good enough" is the whole point of 80/20
- The Hidden Realities of AI Adoption
- John Cutler's Post
- Your AI strategy has a trust problem, not a tooling problem
- Responsible AI Is Becoming a Growth Strategy