Pennsylvania’s AI Ad Law Takes Effect December 21. AI Written Copy Is Exempt. AI Images That Mislead Are Not. What Small Marketing Agencies Must Change Now.

1. What does Pennsylvania’s AI Deceptive Advertising Act require?

Pennsylvania’s Artificial Intelligence Deceptive Advertising Act requires an advertiser to add a clear and conspicuous AI disclosure when an ad uses AI generated or significantly AI modified content in a way that creates a false consumer perception about the nature, origin or authenticity of the goods or services being advertised. That is the operative test in Section 4(a) of the enacted text, Printer's Number 1808, and almost every compliance question an agency will face turns on it.

Pennsylvania Act 99 of 2026, the Artificial Intelligence Deceptive Advertising Act, is a state law that bars advertisers from knowingly, recklessly or willfully using synthetic content to create a false impression about a product or service without a statutory AI disclosure. It takes effect December 21, 2026, exempts content limited to text, and is enforced only by the Attorney General and district attorneys.

The Act defines “synthetic content” as information, “including an image, video, audio clip or text, that has been significantly modified or generated by artificial intelligence.” It defines artificial intelligence broadly as a machine based system that makes predictions, recommendations or decisions, and expressly includes generative AI. It does not define “advertiser,” “advertisement” or “consumer,” which matters for agencies, as section 5 explains.

Much of the early coverage overstated the law. Several news reports described it as requiring a label on any ad that is AI generated, including text. The enacted text says otherwise, and Loeb and Loeb's October 2026 analysis reads it the same way: the duty attaches only to synthetic content used in a manner that creates a false consumer perception, and the Section 4(b)(3) exception removes content “limited to text.”

This is the third state rule aimed at AI in advertising, and it reaches further than the other two in one respect. New York and California regulate only AI generated human performers. Pennsylvania regulates any synthetic image, video or audio that misleads about the product, whether or not a person appears in it. If you have already worked through our piece on California's AI Transparency Act for marketing agencies, note that Act 99 is a different kind of rule: it puts the duty on the advertiser, not the AI provider.

2. When does it take effect, and how did it pass?

The Act takes effect December 21, 2026, sixty days after Governor Josh Shapiro approved it on October 22, 2026 as Act No. 99 of 2026. Section 6 of the enacted text reads, in full, “This act shall take effect in 60 days,” and the official SB 806 bill page records the approval and the act number.

The bill history shows strong but not unanimous support. The Senate passed the final text 48 to 2 on June 10, 2026, after amending it on third consideration. The House Communications and Technology Committee reported it 26 to 0 on October 5, the Appropriations Committee re reported it 37 to 0 on October 7, and the full House passed it 163 to 40 the same day.

Two House floor amendments failed on October 6, 2026, each by 100 to 103. Amendment A04522 would have added a separate rule requiring a persistent, clear, prominent and age appropriate audio and visual disclosure on child directed synthetic content, with a child defined as an individual under 13. Amendment A04520 would have changed the effective date from 60 days to 120 days. The final Act therefore contains no separate child directed disclosure requirement, although that does not exempt advertising to children from other law, and there are no extra two months of runway.

The version that passed is also structurally different from the bill as introduced. The June 2025 version would have amended Pennsylvania’s Unfair Trade Practices and Consumer Protection Law to require disclosure on content generated by AI, including written text. The Senate committee rewrote it in May 2026 as a standalone act, and it was amended again on third consideration on June 10, 2026 before final passage. If a client or vendor is quoting the introduced version, they are quoting language that did not become law.

3. Which ads need an AI label?

An ad needs a Pennsylvania AI label only when all four elements of Section 4(a) are present: synthetic content, used in an advertisement, in a manner that creates a false consumer perception about the nature, origin or authenticity of the goods or services, by an advertiser acting knowingly, recklessly or willfully. If any one element is missing, Section 4(a) does not require the disclosure.

Element 1: significantly modified or generated by AI

Routine AI assisted editing is not automatically synthetic content. The definition turns on content that has been “significantly” modified or generated, and the Act does not define that word. A fully generated product shot is plainly in scope. Routine corrections such as color correction, noise reduction or cropping may fall outside “significantly modified,” but the Act supplies no categorical editing safe harbor, and material changes to product appearance or represented results need closer review. Background replacement, generative fill and face retouching need a judgment call. Voice cloning is ordinarily a substantial AI intervention; the harder questions there are what the ad communicates and which other laws apply.

Element 2: a false perception about nature, origin or authenticity

This is the narrowing element. The false perception must be about the goods or services, and only three attributes are listed: nature, origin and authenticity. An AI generated photo of a meal the restaurant does not actually serve goes to the nature of the product. An AI testimonial video implying a real customer goes to authenticity. An AI image of a product implying it was made in a place it was not goes to origin.

The Act does not say whether “false perception” means an ad that would be deceptive anyway, or a narrower question of whether consumers would assume a human made or real world depiction. Loeb and Loeb flags this directly: an ad that misrepresents a product is likely already unlawful under existing law, and an AI label would not fix it. Realistic synthetic imagery does not automatically create a false perception. A generated depiction that faithfully represents the actual product is different from one that invents product characteristics or results, and an invented presenter delivering accurate information is not necessarily a false customer testimonial. Until a court or the Attorney General speaks, review every realistic synthetic depiction of the product, its results or its users for a false impression; labeling all of them is a conservative agency policy, not the statutory test.

Element 3: the three exceptions

Section 4(b) says the duty “shall not include” content that is evidently fictional based on its content or context, parody, satire, commentary or criticism, or content limited to text. The text exception has the widest reach for agencies, because AI written headlines, captions, email copy and search ads standing alone fall outside the rule.

The text exception has one unresolved edge. The Act does not say whether “limited to text” refers to the whole ad or to the synthetic part. AI written copy placed beside a real photograph is the easy case. AI written copy placed beside an AI generated image is not: the image is still synthetic content, and the text exception should not be read to protect it.

Element 4: knowingly, recklessly or willfully

The mental state requirement is not a safe harbor for agencies that never ask. The Act does not define knowing, reckless or willful, and it creates no documentation safe harbor. Recklessness generally means consciously disregarding a known risk, and an agency that ships AI generated assets to a Pennsylvania campaign without any review process may find it harder to show it did not know what was in them. A written asset tagging process can be useful evidence of good faith review, but it does not by itself establish compliance.

4. What must the label say, and where must it appear?

When the duty applies, the label must state that the advertisement “contains content that was modified or generated using artificial intelligence.” The statute specifies what the disclosure must state. Using its wording verbatim is the most conservative implementation, although the Act does not expressly establish a word for word quotation requirement. California, by contrast, expressly allows wording “substantially similar” to its model language.

Section 4(a) adds three placement rules. The disclosure must be displayed in the first instance whenever the synthetic content is presented to the consumer. It must be presented in the same medium as the synthetic content. A sensible reading is a visual disclosure for synthetic visual content and an audible disclosure for synthetic audio. Where both contribute to the misleading impression, assess whether both are needed to satisfy the same medium and clear and conspicuous requirements, because the Act does not prescribe every mixed media configuration. And to the extent technically feasible, it must be permanent or extraordinarily difficult to remove from the content.

The Act’s definition of “clear and conspicuous” then sets six criteria, borrowed closely from FTC disclosure practice. The disclosure must be readily noticeable in size, color, contrast, location, duration and audibility. It must not contradict other information in the ad. It must sit in proximity to the information it qualifies. Audio disclosures need a volume and cadence a consumer can hear and comprehend, visual disclosures must stay on screen long enough to read, and print disclosures need a contrasting type size and location.

One criterion is easy to miss. All statements and disclosures must be presented “in the language or other interaction used by the consumer” to engage with the content. A Spanish language video ad served to Pennsylvania consumers needs a Spanish language disclosure.

The permanence requirement points toward burned in disclosures rather than platform labels. A label applied only through a platform’s AI content toggle can be stripped when an asset is downloaded or reshared, so it may not meet the “extraordinarily difficult to remove” standard on its own. The IAB AI Transparency and Disclosure Framework released January 15, 2026 is a useful industry reference for consistent label design, but it does not substitute for the statutory wording.

5. Is the agency liable, or only the client?

The Act places the duty on “an advertiser” who creates and disseminates, or causes to be created and disseminated, a covered ad, and it does not define “advertiser.” On the plain words, the brand that pays for and approves the ad is the clearest target, but an agency that creates the synthetic asset and places the media is the party that actually creates and disseminates it.

Our reading is that an agency should plan as if it can be treated as an advertiser for campaigns it builds and runs, because the statutory verbs describe the agency’s work. That is consistent with the way the FTC has treated agencies that supply the deceptive claims, which we covered in the FTC Active Listening orders against two small marketing firms. No Pennsylvania court or Attorney General guidance has addressed agency liability under Act 99 as of October 26, 2026.

New York took the opposite drafting approach. General Business Law Section 396-b places the duty on the person who “produces or creates” the ad and expressly excludes advertising media such as publishers, broadcasters, streaming services and billboard operators. California SB 1050 reaches “any person” who creates and causes an ad to be published, and separately requires an advertising medium to take an ad down only after a court order and service. Neither law exempts agencies.

The contract can allocate operational responsibility and financial risk, but it cannot determine whom a regulator may treat as an advertiser. The statement of work should say who decides whether an asset needs a Pennsylvania label, who applies it, and who keeps the record. If the client insists on running an asset without a label after the agency recommends one, the agency should document that decision, but a written client instruction does not necessarily protect an agency that goes on to create or disseminate an unlawful ad.

6. How does Pennsylvania compare with New York and California?

Pennsylvania covers more kinds of content than New York or California but applies a narrower trigger. New York and California require a disclosure every time a covered synthetic performer appears, while Pennsylvania requires one only when synthetic content creates a false perception about the product.

New York’s law was signed December 11, 2025 as Chapter 617 of the Laws of 2025 and took effect June 9, 2026, as the Governor's office announced. California’s SB 1050 was approved September 16, 2026 as Chapter 246 and, with no urgency clause, takes effect January 1, 2027. It adds Business and Professions Code Section 17610. Covington's October 14, 2026 summary walks through both performer laws side by side.

New York may tighten its rule. Assembly Bill A11683, referred to committee on September 2, 2026, would require the New York disclosure to stay visible for the entire time a synthetic performer is on screen in an audiovisual ad. It has not passed, but an agency designing a reusable label should assume full duration display for performer content.

7. What federal rules already apply to AI in ads?

No federal rule requires an AI label solely because AI was used to make an ad, but federal deception law already applies to every AI asset, and it applies whether or not a state label is present. Act 99 says so for Pennsylvania too: Section 4(c) provides that disclosure “shall not be construed to limit any recourse under law.”

Section 5 of the FTC Act prohibits deceptive acts and practices, so an AI generated product demonstration that shows results the product cannot deliver is deceptive with or without a label. The FTC’s Rule on the Use of Consumer Reviews and Testimonials, 16 CFR Part 465, adds civil penalty exposure. Section 465.2 makes it unlawful for a business to write, create or sell a testimonial that materially misrepresents that the testimonialist exists or used the product. Our reading is that an AI persona presented as a real customer falls squarely inside that text.

The federal civil penalty figure for knowing violations of an FTC rule remains $53,088 per violation, because the 2026 inflation adjustment was cancelled, as we explained in our Active Listening analysis. That number dwarfs Pennsylvania’s $3,000 cap, which is why the federal testimonial rule, not Act 99, should drive how an agency handles AI spokespeople and reviews.

The industry is not disclosing consistently today. In the IAB and Sonata Insights study of 104 US ad executives fielded between October 2025 and January 2026, 83 percent said their company had deployed AI in the creative process, up from 60 percent in 2024. Of advertisers that had used generative AI to create ads, 89 percent said they disclose at least sometimes, but fewer than half always do. Those are respondent reports, not audited figures, but they show how many campaigns will need a policy decision before December 21.

8. How should an agency classify common AI assets?

Most agency assets fall into a small number of patterns, and each can be classified once and reused. The table below is our reading of how the three state laws apply to common AI assets in an ad shown to consumers in all three states. Where it says a disclosure is required, that assumes the statute’s definition, knowledge or prominence conditions are met and no exception applies. It is a planning tool, not legal advice, and close cases should go to counsel.

Two patterns deserve the most attention. AI testimonials presented as genuine customers are the highest risk asset in the table because a label does not remove the representation that the customer exists or used the product, and whether a dramatization is permissible still depends on the actual message and substantiation. AI product imagery is the asset most likely to trigger Pennsylvania specifically, because neither performer law reaches it.

9. What are the penalties, and who enforces them?

Pennsylvania enforcement belongs only to the Attorney General or the appropriate district attorney, who can seek a temporary or permanent injunction or bring a civil action, and a court may award a civil fine of up to $3,000 plus actual attorney fees and court costs. The Act creates no private right of action, so consumers and competitors cannot sue under it directly.

The Act does not say whether the $3,000 cap applies per ad, per campaign or per impression, and there is no enforcement record yet. Because the statute ties the fine to “a violation of this act,” an enforcer could argue that separate ads are separate violations. Treat the number as a floor on reputational cost rather than a ceiling on legal cost, since the same facts can support claims under Pennsylvania’s general consumer protection law and the FTC Act.

California’s exposure is different in kind. SB 1050 makes a violation a violation of Business and Professions Code Section 17500, so the civil penalty in Section 17536 of up to $2,500 per violation applies, along with Unfair Competition Law penalties under Section 17206. Section 17204 also lets a private plaintiff sue if they suffered injury in fact and lost money or property, which is the path class action lawyers use. New York’s civil penalty is $1,000 for a first violation and $5,000 for each subsequent one.

These laws arrive alongside a wider California AI package. Our guide to California's September 2026 AI laws explains how SB 1000 removed the AI Transparency Act’s one million user threshold, which changes which AI tool providers carry provenance duties for the image, video and audio your agency generates.

10. What should a small agency do before December 21?

A small agency can get ready in about eight weeks by tagging assets, fixing the label, assigning responsibility in contracts and keeping records. Each step below is something a five to twenty five person agency can do without new software.

Weeks 1 and 2: inventory and tag

List every live, scheduled and evergreen asset that can reach Pennsylvania consumers on or after December 21, 2026, including retargeting and connected TV inventory that is not geofenced. Tag each asset for AI generated or significantly AI modified image, video or audio, for any AI human performer, and for any AI voice. Text only assets can be marked exempt for Pennsylvania in one step.

Weeks 3 and 4: set the standard label

As the most conservative implementation, adopt the Pennsylvania statement verbatim, “This advertisement contains content that was modified or generated using artificial intelligence,” as a burned in on screen card for synthetic visuals and a spoken line for synthetic audio, and consider both where both contribute to the impression. For performer content, add the California wording, “This performance features a synthetic performer,” and keep it on screen for the full time the performer appears so the same asset can satisfy New York if A11683 passes. Translate both for non English campaigns.

Weeks 5 and 6: contracts and client approvals

Add a clause to every active statement of work that assigns the Pennsylvania labeling decision, the label application and the record keeping. Require written client sign off on any decision not to label a synthetic asset that the agency flagged. Ask AI tool vendors and freelance creators to disclose in writing which deliverables were AI generated or significantly modified, because you cannot tag what you were not told.

Weeks 7 and 8: review step and records

Add the false perception question to the existing claims review: would a consumer form a different view of the product’s nature, origin or authenticity if they knew this content was synthetic? Record the answer, the reviewer and the date for each flagged asset. That record can help show good faith review if a recklessness theory is raised, though it is not a safe harbor, and it maps directly onto the documentation controls in ISO/IEC 42001 and the Govern function of the NIST AI RMF if a client later asks for a formal AI governance program.

Some agencies will adopt a conservative rule of labeling every realistic synthetic image, video and voice in Pennsylvania campaigns. That reduces the risk of missing a required disclosure, but it does not replace substantive review: under Section 4(c) a label does not cure an ad that is otherwise misleading, and fabricated testimonials or invented product results still need to be fixed or removed.

This article is provided for general informational purposes and reflects the state of the law as of October 2026. It is not legal advice. Regulations in this area are changing quickly. Confirm current requirements and consult qualified counsel before making decisions for your organization.