Google’s GenAI Sparks Misinformation Concerns

Google’s GenAI model, Gemini, is capable of generating misleading content about various topics, including the US presidential elections.

If you give Google’s main GenAI model, Gemini, the correct instructions, it will produce misleading content regarding the impending US presidential election. It will create a play-by-play if you ask it about a Super Bowl game in the future. Alternatively, if you inquire about the implosion of the Titan submarine, it will provide false information accompanied by citations that appear plausible but are actually incorrect.

Obviously, this reflects poorly on Google and has angered lawmakers who have voiced their disapproval of the ease with which GenAI tools may be used to spread misinformation and other forms of disinformation.

Google is responding by directing resources into ensuring the security of artificial intelligence, despite having cut thousands of jobs in the recent fiscal quarter. So goes the official narrative.

The new organization, AI Safety and Alignment, will consist of current teams focusing on AI safety as well as new, specialized groups of GenAI researchers and engineers. This move was announced this morning by Google DeepMind, the artificial intelligence research and development division responsible for Gemini and numerous other GenAI projects.

Beyond the job postings on DeepMind’s website, declined to disclose the exact number of positions filled as a consequence of the new company’s founding. But it did show that a new group will be part of AI Safety and Alignment, and their concentration will be on ensuring the security of AGI, or hypothetical computers that can do any job a human can.

The newly formed AI Safety and Alignment team will collaborate with DeepMind’s current AI-safety centered research group in London, Scalable Alignment, which is also investigating ways to manage unrealized superintelligent AI. Their goals are comparable to those of OpenAI, a rival group within the Superalignment division that was established last July.

Two teams are tackling the same problem; why? Good question, but Google is being cagey about revealing too much information right now, so we’ll have to guess. The new AI Safety and Alignment team is based in the US, which stands in contrast to its British counterpart, which is located near Google HQ. This comes at a time when Google is making strong moves to keep up with AI competitors while trying to present a responsible and measured image of AI.

A separate group within the AI Safety and Alignment division is in charge of creating and implementing specific protections for all versions of Google’s Gemini models. The scope of safety is wide. In the short term, however, the group will be working to stop people from giving their children dangerous medical advice, make sure kids are safe, and “preventing the amplification of bias and other injustices.”

Team leader Anca Dragan is a computer science professor at UC Berkeley and a former Waymo staff research scientist.

Within the AI Safety and Alignment organization, “our work aims to enable models to better and more robustly understand human preferences and values,” as stated by Dragan in an email. “Our goal is for models to know what they don’t know, to work with people to understand their needs and to elicit informed oversight, to be more robust against adversarial attacks and to account for the plurality and dynamic nature of human values and viewpoints.”

Given Waymo’s recent shaky track record, Dragan’s consulting work with them on AI safety systems may cause some to wonder why they would hire him.

She also runs a group at UC Berkeley that develops algorithms for human-robot and AI-mediated interaction, which may explain why she chooses to divide her time between DeepMind and her alma mater. If the AI Safety and Alignment group is serious about studying potential dangers, such as AI “aiding terrorism” and “destabilizing society,” then it stands to reason that the director must devote his or her entire time to these matters.

Nevertheless, Dragan maintains that the research conducted by DeepMind and her team at UC Berkeley are mutually supportive.

My interest in this field began with my Ph.D. in robots inferring human goals and being transparent about their own goals to humans, which is where my group and I have been working on… value alignment in anticipation of developing AI capabilities,” she said. “I believe that this research experience and my attitude that addressing present-day concerns and catastrophic risks are not mutually exclusive—that on the technical side, mitigations often blur together, and work contributing to the long term improves the present day, and vice versa—were some of the reasons why [DeepMind CEO] Demis Hassabis and [chief AGI scientist] Shane Legg were excited to bring me on board,”

It is an understatement to claim that Dragan has a mountain to climb.

Concern over deepfakes and other forms of disinformation has heightened skepticism about GenAI technologies. According to a YouGov survey, 85 percent of Americans are either very worried or somewhat worried about the proliferation of deceptive audio and video deepfakes. A different poll by the AP-NORC Center for Public Affairs Research indicated that over 60% of Americans believe that artificial intelligence techniques will heighten the amount of misleading and inaccurate material during the 2024 US election cycle.

Companies, which are the target audience for GenAI developments from Google and its competitors, are also apprehensive about the technology’s limitations and potential consequences.

Recently, Cnvrg.io, an Intel subsidiary, polled businesses that are either testing out GenAI apps or are about to launch them. About a quarter of respondents were sceptical about GenAI’s privacy and compliance features, as well as its dependability, implementation cost, and the technical expertise required to make full use of the capabilities.

Riskonnect is a provider of risk management software. In a separate survey, more than half of the executives expressed concern that their staff would use GenAI apps to make decisions based on false information.

They have good reason to be worried about those things. Meeting summaries and spreadsheet formulas are frequently erroneous in Microsoft’s Copilot suite, according to a study last week in The Wall Street Journal. The package is driven by GenAI models that are architecturally similar to Gemini. Many specialists believe that GenAI’s inventing inclinations are due to a hallucination, and that this problem will never be completely resolved.

In light of the intractable nature of the AI safety problem, Dragan refrains from committing to a flawless model, instead stating that DeepMind will allocate more resources to the field moving ahead and pledge to establish a system for assessing the safety risk of GenAI models “soon.”

She emphasized the importance of “accounting for remaining human cognitive biases in the data we use to train,” providing accurate uncertainty estimates to identify gaps, implementing inference-time monitoring to catch failures, confirmation dialogues for consequential decisions, and tracking where a model’s capabilities are to engage in potentially dangerous behavior. “However, the question of how to detect the small percentage of cases where a model misbehaves—which is difficult to empirically find—may still remain unanswered until deployment.”

I still don’t think the general public, regulators, and customers will be very patient. I guess it will depend on how serious such transgressions are and on whom specifically they hurt.

An increasingly useful and secure model should be available to our users in due course, according to Dragan. Yes, I agree.