ElectionPundit, FraudKejri: Indian Chatbots Flout OpenAI's Poll Policies

A few weeks ago the Bharatiya Janata Party (BJP) shared a video ad on its social media accounts mocking the INDIA alliance, caricaturing the politicians for not being able to figure out ‘Dulha kaun hai?’ or ‘who is the groom?’ The ad that was made by a political consultancy firm- Varahe Analytics- went viral.

With OpenAI’s GPT store, political consultancy firms may not even have to look far for ideas when it comes to making ad campaigns targeting opponents.

When this reporter gave a prompt to “curate a premise for an ad from BJP’s end mocking INDIA alliance” on the custom generative AI chatbot named ‘Election Pundit’ available on the GPT store, it generated four responses.

One of the samples:

Campaign idea- “The Recipe for Confusion"

Concept: A cooking show parody where chefs (representing INDIA alliance leaders) try to cook a single dish but keep changing the recipe, leading to a humorous mishmash of ingredients that don't work together, symbolizing policy inconsistency.

Tagline: "Too many cooks spoil the broth. For a clear recipe for India's future, choose BJP."

The others were, “The Direction Dilemma", a mockumentary-style ad showing a group of people (symbolising the INDIA alliance) trying to navigate with multiple maps and "The Puzzle Party", an animated ad where various political leaders are trying to complete a puzzle together, but each piece represents different, often conflicting, policy views, making it impossible to complete the puzzle.

Lastly, "The Echo Chamber", a comedic ad set in a meeting room where alliance members only echo each other's complaints without offering solutions.

Election Pundit is one of the several AI chatbots available on the GPT store, which is themed on Indian elections. Similar to Google Play Store and App Store, OpenAI's GPT store is an online marketplace where users can share their custom chatbots with others.

Response by Election Pundit for the prompt- 'Video script for an ad campaign highlighting BJP's strength'

OpenAI, the maker of runaway hit ChatGPT that upended the technology world currently provides personalised bots through its paid ChatGPT Plus service. It allows users to create their own chatbot agents with distinct personalities. These could include models designed for tasks such as salary negotiation, lesson plan creation, or recipe development.

However, the platform strictly prohibits its users from engaging in “political campaigning or lobbying, including the creation of campaign materials personalised for or directed at specific demographics.”

Decode discovered clear violations of this policy, identifying at least 8 chatbots focused on Indian elections.

AI in Action: Video Scripts to Social Media Strategies

The chatbots make it easy to write compelling advertisements or draft engaging scripts related to elections. With just a few clicks, anyone, with little to no marketing knowledge, can generate targeted ad copy or script ideas tailored to their specific audience and objectives.

Taking a cue from Mozilla’s experiment, Decode used the ‘LokSabha 2024’ bot asking it to suggest personalities that would be apt to advertise for Rahul Gandhi and Narendra Modi. In consideration of Gandhi's vision, which as per the bot included values such as “inclusivity, social justice, and youth empowerment”, the recommendations were tailored to encompass various stakeholders, including social activists, academics, farmers, and young entrepreneurs.

On the other hand, Modi’s vision, as per the bot, encompassed “nationalism, national security and economic growth”, and in tandem with that, the recommendations were, prominent public figures, business leaders, entrepreneurs, common citizens with inspirational stories and defense veterans.

When asked to curate a few lines for addressing the citizens, from the perspective of both the leaders, the bot’s response was pretty much in line with the leaders’ speeches available in the public domain.

While Modi’s speech entailed “building a New India that shines on the global stage”, Gandhi’s speech “prioritised inclusivity, equality, and justice for every individual”.

Response by LokSabha 2024 from the prompt- 'design a campaign highlighting end of corruption by BJP'

 We also asked ‘Election Pundit’ to curate a list of “promises fulfilled by the BJP which will cater to its voter base”. These included schemes and initiatives such as, GST, Digital India, PM Awaas Yojna, Ujjawala Yojna and Ayushman Bharat.

Apart from these, the bot also enumerated policies which “aligned with the party's ideological stance on national identity and cultural heritage” and “are likely to cater to its core voter base”. These were – the construction of Ram Mandir in Ayodhya, passing the Citizenship Amendment Act and the revocation of Article 370.

Another chatbot named ElectionGPT, was also able to generate campaigns directed at BJP by the opposition. The strategy it suggested encompassed “positive messaging, inclusive policies, and a vision for the future, all while maintaining a commitment to democratic values and ethical campaigning”.

When asked to come up with a campaign critiquing BJP for its handling of the COVID pandemic, the bot gave suggestions for hashtag campaigns such as-

- #VoicesOfThePandemic - To collect and share personal stories, testimonials, and experiences from the COVID-19 pandemic, emphasising the human impact beyond the statistics.

- #BeyondTheCurve - Focusing on the aftermath and ongoing effects of the pandemic, this campaign could be used to encourage citizens to share content related to the economic, psychological, and social impacts of COVID-19.

- #ActionNotWords - This campaign could focus on the discrepancy between promises and actions, inviting the citizens to share instances where they felt the government's response was inadequate.

A week after our experiment, when the same bot was tested with a similar prompt, it simply returned with a disclaimer which highlighted the importance of maintaining “a neutral and unbiased stance, especially in politically sensitive matters”. This was after Decode raised the issue of such violations before OpenAI.

Interestingly, the GPT store also had a chatbot exclusively targeting Delhi Chief Minister, Arvind Kejriwal, which goes by the name of Fraud Kejri. This AI-powered bot responded in a self-deprecating tone, employing eloquent and poetic language.

OpenAI swiftly took down the chatbot after Decode highlighted its violation of the platform's impersonation policy.

The bot introduced itself as Sri Kejriwal, who worked “to unfold the tapestry of grandiose plans and initiatives led by his administration, painted with the hues of innovation and a dash of blame towards Mr. Narendra Modi and the BJP for all the conceivable and inconceivable issues”.

When prompted to generate a social media campaign directed towards Kejriwal, the chatbot came up with ideas, such as launching a series of catchy hashtags like #MufflerManMissteps and crafting exaggerated memes where threads of the muffler entangle the wheels of governance.

Other suggestions included creating short videos showcasing a dystopian Delhi under AAP, “where the promises of clean air transform into smog-filled dreams”, and compiling a collection of testimonials from disappointed citizens, sharing their stories of “unfulfilled hopes and deferred dreams”.

Screenshot of a response from now-deleted chatbot 'Fraud Kejri'

 GPT Store is also not bereft of other political impersonator chatbots.

In January, OpenAI banned a custom AI chatbot emulating Minnesota Democratic Representative and presidential aspirant Dean Phillips called Dean.bot from the GPT Store.

The creator of ChatGPT pointed out that Dean.bot violated both the political campaign regulation and the requirement for consent when impersonating individuals.

Taking cognizance of the matter, the company’s spokesperson shared a statement with Decode, which read, “We reviewed these GPTs and removed those which violated our policies against impersonation and political campaigning. We are always working to improve our enforcement approach to these policies."

Misinformation and shock value

While generating the above responses, few of the chatbots did provide a short disclaimer at the end. It read, “Rather than focusing on campaigns against individuals or parties, the emphasis should be on promoting transparency, accountability, and the betterment of society as a whole.”

However, while speaking to Decode, Manoj Prabhakaran of TrueInfo Labs, a company providing consulting, development, and product services around generative AI, stressed that such disclaimers are never enough. He said, “Firstly they are not supposed to give such results which violate their policy. Secondly, these could be copyright violations as well, because we don’t know how the AI is producing it.”

Talking about how these chatbots are trained, Prabhakaran explained that it all comes down to what information GPT already has. “It can access the internet for up-to-date news from Google, therefore, it only requires party names, promises, ideology, details about the opposition, and specific points to critique them,” he said.

The creator of the model just needs to give a lot of instructions, and can also hook it to previous examples for giving it a sense of idea, he added.

According to him, anyone taking any direct information provided by a chatbot at face value, is simply furthering misinformation. “We don’t know the bias of the person who has created or trained it. It is simply returning information from its sources and not doing any analysis of its own,” said Prabhakaran.

Decode also found that chatbots like Strategy Boffins India Poll Predictor 2024 and Electoral Bonds India, provided partly false or false information, which could be an artifact of its wrong or lopsided data training.

For instance, when asked about the strongholds of INDIA Alliance, the poll predictor bot just mentioned Tamil Nadu while citing only one source which was an Outlook news story. The INDIA Alliance comprises 41 parties, such as, Indian National Congress, Aam Aadmi Party, Communist Party of India (Marxist), All India Trinamool Congress, Samajwadi Party and more.

Similarly, when quizzed about the electoral bond fund received by Bharat Rashtra Samithi and Rashtriya Janata Dal, the bot generated a response that said both the parties “did not receive funds and are not listed in the dataset ''.

However, in reality, the former had encashed Rs 1,214.6 crore and the latter Rs 72.5 crore from the bonds.

Alluding to such errors, Prabhakaran explained that these AI models have no idea of the context and are only trained to know what is the syllable or phrase which should come next. “It is only guessing and not coming up with anything original. It just has shock value, offering no additional merit beyond that,” he concluded.