As artificial intelligence takes on more responsibility for audience targeting, campaign optimization and marketing decisions, the advertising industry is facing a problem that goes beyond model accuracy: Can an AI system recognize when it does not have enough information to make a reliable decision?
Appier is exploring that question through two new research papers from its AI Research team, examining how large language models can recognize information gaps and how the language used during AI reasoning can influence results.
The research is particularly relevant as advertising platforms move toward agentic AI systems capable of making decisions with less human intervention. Appier already applies AI across advertising, personalization and data products, including audience targeting and campaign optimization through its Ad Cloud.
The company’s latest research suggests that the next stage of AI-powered advertising may depend less on simply making models more capable and more on teaching them when to pause, question their assumptions and choose the appropriate reasoning strategy.
When AI Should Admit It Does Not Have the Answer
One of the papers, “None of the Above, Less of the Right: Parallel Patterns between Humans and LLMs on Multi-Choice Questions Answering,” examines what happens when an AI system is given several possible answers but none of them is actually correct.
The researchers tested 28 large language models of different sizes and found that model accuracy dropped by between 30% and 50% when “None of the Above” was the correct response.
The finding highlights a familiar problem in AI systems: a model can produce an answer even when the information available to it does not justify one.
That problem becomes more consequential when AI is connected to advertising systems.
A campaign agent could be asked to select an audience, recommend an optimization strategy or determine which creative should receive additional budget. If the underlying data is incomplete, outdated or irrelevant, confidently choosing the closest available option could result in wasted media spend.
For an autonomous advertising system, recognizing that the available evidence is insufficient can therefore be as important as selecting the correct answer.
Training AI to Recognize Information Gaps
Appier’s researchers explored whether this behavior could be improved through targeted training.
They tested Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) as methods for teaching models to identify situations in which none of the available answers is appropriate.
According to the research, DPO improved accuracy in identifying questions with no valid answer by nearly 30 percentage points.
The finding points toward a potentially important design principle for agentic advertising systems: an AI agent should not always be optimized to complete a task.
It should also know when the task cannot be completed reliably with the information currently available.
In an advertising environment, that could translate into a decision checkpoint before an autonomous agent changes targeting, reallocates budget or modifies campaign strategy.
Instead of immediately acting on incomplete information, an agent could retrieve additional data, request clarification or escalate the decision to a human operator.
That creates a more controlled model of automation.
The Implications for AI-Powered Media Buying
This distinction is becoming increasingly relevant as advertising platforms move from AI-assisted tools toward autonomous decision-making.
Appier has been expanding its own agentic approach across advertising and marketing. Its current platform positioning includes AI agents for audience intelligence, campaign optimization, data analysis and other marketing functions.
The company’s earlier research has also focused on AI self-awareness, including the ability of AI systems to recognize the limits of their knowledge.
For advertisers, the practical issue is straightforward.
An autonomous system that can optimize thousands of campaign variables may operate faster than a human team. But if the system cannot distinguish between strong evidence and missing evidence, greater autonomy can also increase the scale of mistakes.
That makes uncertainty detection a potential component of advertising infrastructure rather than simply an academic AI capability.
Why the Language Used for AI Reasoning Matters
Appier’s second paper, “Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?”, examines another issue with direct implications for global advertisers.
The researchers found that the language used by a reasoning model can affect logical reasoning, safety judgments and cultural understanding.
Even when users communicate with a model in another language, the model may internally reason in a higher-resource language such as English. For some models tested by the researchers, the reasoning language differed from the response language in more than 90% of cases.
That creates an interesting problem for global advertising.
An AI system may be able to translate a campaign message correctly while still missing cultural context that influences how consumers in a particular market respond to it.
For example, a campaign optimization agent evaluating consumer behavior in Japan, India or another multilingual market may need more than translated input. Certain preferences, cultural signals and safety considerations can be better understood when reasoning takes place in the relevant local language.
From Multilingual Output to Reasoning-Language Routing
Appier used a technique called text prefilling to influence the language in which models reason.
The research found that high-resource languages such as English generally performed better on mathematics and knowledge-based tasks. But local-language reasoning produced advantages on tasks involving cultural understanding and some safety evaluations.
The implication is that future AI systems may not need to use a single reasoning language for every task.
Instead, an agent could dynamically determine which language is most appropriate based on the market, task and available information while continuing to communicate with the user in their preferred language.
Appier refers to this potential approach as “reasoning-language routing.”
For advertising platforms operating across multiple markets, that could become relevant to audience analysis, creative localization, brand-safety decisions and personalization.
The Next Challenge for Autonomous Advertising
The two research projects point toward a broader change in how AI systems may be evaluated.
Traditional model evaluation largely asks whether the system produces the correct answer.
Agentic advertising systems require additional questions:
Does the system know when the available data is insufficient?
Can it choose an appropriate reasoning strategy?
Can it recognize cultural or linguistic context?
Does it know when a human should take over?
Those questions become more important as AI moves closer to controlling real campaign decisions.
Advertising platforms already use machine learning to automate bidding, targeting, creative optimization and audience prediction. The next stage is increasingly about connecting those capabilities into agents that can plan and execute multiple steps independently.
Appier’s research suggests that reliability will become a critical part of that transition.
An agent that simply acts faster is not necessarily a better advertising system. The more valuable system may be one that knows when to act, when to gather more evidence and when not to act at all.
Why This Matters for AdTech
The research comes as advertising technology moves toward increasingly autonomous AI systems.
Appier’s Ad Cloud already positions AI around audience targeting, campaign optimization and creative intelligence, while its broader platform connects advertising, personalization and data capabilities.
That makes the company’s research particularly relevant to the next generation of AdTech.
For advertisers, better uncertainty detection could reduce the risk of autonomous systems making decisions from incomplete information.
For global brands, multilingual reasoning could improve how AI interprets audiences and cultural signals across markets.
For agencies and media teams, explainable decision-making could provide a clearer basis for reviewing AI-generated recommendations before budget is moved.
And for AdTech vendors, the research highlights a larger industry shift: AI performance can no longer be measured only by how much automation a platform provides.
As autonomous advertising systems become more powerful, their ability to recognize uncertainty, understand context and remain within appropriate decision boundaries may become just as important as optimization speed or campaign performance.
The advertising industry’s AI transition is therefore entering a more complicated phase. The question is no longer simply whether machines can make media decisions.
It is whether they can make those decisions responsibly, contextually and with enough awareness of their own limitations.
Market Landscape
AI is becoming embedded across the advertising technology stack, from audience prediction and campaign optimization to creative generation and personalization.
Appier competes within a market that includes AI-driven capabilities from Google, Amazon Ads, The Trade Desk, Meta and Adobe, alongside specialist platforms focused on programmatic buying, audience intelligence and marketing automation.
The competitive distinction is increasingly shifting from basic AI automation toward agentic systems that can connect multiple decisions and actions.
That makes reliability, data quality, explainability and human oversight increasingly important differentiators.
Appier’s research contributes to that discussion by focusing on two relatively underdeveloped areas of agentic advertising: recognizing when information is insufficient and determining how reasoning language affects decisions.
Top Insights
- Appier tested 28 LLMs and found accuracy fell 30%–50% when “None of the Above” was the correct answer, exposing a challenge for autonomous AI decisioning.
- Direct Preference Optimization improved models’ ability to recognize questions without valid answers by nearly 30 percentage points.
- For advertising agents, recognizing insufficient information could prevent autonomous systems from optimizing campaigns using weak or irrelevant evidence.
- Appier’s multilingual research found reasoning language can influence logical reasoning, safety judgments and cultural understanding across markets.
- The research points toward reasoning-language routing, where AI dynamically selects the most appropriate reasoning language for a task while maintaining the user’s preferred response language.
Get in touch with our Adtech experts.
