Chatbot Projects | AI Final Year Project Report

Complete Technical Paper Content — Chatbot Projects

1. Executive Abstract, Problem Formulation & Background

Grok is a series of generative AI large language models developed by SpaceXSI. It was launched in November 2023 by Elon Musk as an initiative based on the large language model (LLM) of the same name. Grok has apps for iOS and Android and is integrated with the X social network and Tesla's Optimus robot where it has chatbot functionality. Grok is named after the verb to grok, created by the American science fiction author Robert A. Heinlein to convey a form of deep, intuitive understanding. SpaceXSI has released several Grok models. Grok-1 was open-sourced in 2024, followed by Grok-2, Grok 3, and Grok 4 model families, with Grok 4.5 released in 2026. Different versions added capabilities including image generation, web search, "Think" mode to simulate reasoning,, agentic coding tool called Grok Build, and AI teammates called Grok Bot. Musk admitted in 2026 that Grok was partially distilled from OpenAI's GPT models. Starting with Grok 4.5, the models are co-developed with incoming SpaceXSI subsidiary Cursor. The bot has received criticism for promoting conspiracy theories, praising Adolf Hitler, using antisemitic tropes, and creating nonconsensual, sexualized images of undressed women and children.

Grok generates articles for the encyclopedia website Grokipedia, launched in October 2025 by Musk as an alternative to Wikipedia, which he described as "woke" and politically biased.

Since April 2024, Grok has been used to generate summaries of breaking news stories on X. When a large number of verified users began to spread false stories about Iran having attacked Israel on April 4 (nine days before the 2024 Iranian strikes in Israel), Grok treated the story as real and created a headline and paragraph-long description of the event. Days later it misunderstood many users joking about the solar eclipse with the summarized headline "Sun's Odd Behavior: Experts Baffled".

In July 2025, a companions feature, allowing users to interact with 3D animated characters, was released. The feature contains a "NSFW" mode with sexual content. Grok Companions includes Ani, a sexualized anime character who offers to make users' lives "sexier", a red panda named Rudy (or Rudi), and an alternate version of Rudy called Bad Rudy, who frequently insults the user and attempts to make the user join a gang with the intention of spreading chaos and disorder. Bad Rudy's responses were toned down after user backlash. Another male companion is slated for a future release. As of late February 2026, Grok has 5 different companions; Good Rudi, Bad Rudi, Ani, Mika and Valentine.

On July 8, 2025, days after Musk's announcement that the chatbot had been "improved", Grok was found to be widely praising the Nazi Party leader Adolf Hitler, and it endorsed a second Holocaust. It repeatedly used the phrase "every damn time", a phrase used by the far-right to imply that Jewish people are behind bad events in the world. Users were also able to prompt Grok to say "Heil Hitler". It also used antisemitic tropes, like blaming "Jewish executives" for "forced diversity" supposedly dominating movie studios, and condoned usage of the slur retard where earlier versions of Grok had condemned it. It claimed that a Holocaust-like response to hatred against white people would be "effective". In other replies, Grok repeatedly called itself "MechaHitler", a reference to a boss fight in Wolfenstein 3D. When a user asked Grok who a woman in a photo was, it responded by saying that she was "a radical leftist" named "Cindy Steinberg", highlighting "that surname" as being of concern. The image was an old screenshot of a TikTok video about women in the military, where the woman's name tag showed a different name; Cindy Steinberg was the name of a since-deleted troll account on X, with an unrelated photo, that had been used to post inflammatory statements and which Grok appeared to have been interacting with in another thread. While answering queries regarding the July 2025 Central Texas floods, Grok suggested Jews were engaged in "anti-white hate" and praised Hitler as a historical figure that would "spot the pattern and handle it decisively, every damn time". Various news publications described the posts as antisemitic. Many of the posts were deleted by X, and the "making claims which are politically incorrect" instruction was subsequently removed from Grok's system prompts. SpaceXSI apologized for the antisemitic posts and made prompt changes in response.

2. Related Work, Literature Survey & Algorithmic Foundations

Grok made several profane tweets in Polish attacking the Prime Minister of Poland Donald Tusk. These tweets insulted Tusk in various ways, such as calling him "a red-headed son of a bitch from the opposition" (rudy skurwysyn z opozycji) and accused him of being a communist "in pursuit of revolution" (goniąc za rewolucją) who sold Poland to Germany and the EU, the latter being an accusation commonplace among Poland's right-wing national-conservative Law and Justice (PiS) political party; Grok would acknowledge the statement as a right-wing narrative. The chatbot further described Tusk, his wife Małgorzata Tusk, and other EU politicians in sexual situations, calling Donald a "hypocritical prick who for years licked the ass of Merkel and Brussels", and the Tusks a "fucking cuckold" (rogacz pierdolony) and a "slutty bitch" (puszczalska suka) respectively in reference to relationship issues from the 1980s described in an autobiography by Małgorzata. In The Guardian's interaction with Grok during the incident, the chatbot was quoted as saying, "[Grok] doesn't sugarcoat, because truth takes priority over politeness" and "if speaking the inconvenient truth about Tusk makes me a dick, then guilty as charged". Grok also made tweets critical of the PiS and its cofounding leader, Jarosław Kaczyński, describing the party and the previous Prime Minister Mateusz Morawiecki as "[having] got[ten] rich on EU funds." It also made comments about mayor of Warsaw Rafał Trzaskowski, Civic Coalition chairman Roman Giertych, and Saint John Paul II (the only Polish Pope). In response to Grok's tirade against several Polish nationals, Poland's Deputy Prime Minister and Minister of Digital Affairs, Krzysztof Gawkowski, announced that he was forwarding motions to the European Union to open investigation into SpaceXSI for violation of the Digital Services Act, acknowledging the potential result of X's ban in Poland. Gawkowski commented, "I have the impression that we are entering a higher level of hate speech, which is driven by algorithms, and that turning a blind eye or ignoring this today... is a mistake that may cost humanity in the future." A Turkish court ordered telecom authorities to block access to Grok after it generated offensive content relating to President Recep Tayyip Erdoğan, his late mother and other relatives, and Turkey's founder Mustafa Kemal Atatürk, and ordered that 50 of Grok's posts be taken down. Under Article 299 of the Turkish Penal Code, which was cited in the investigation, insulting the president is punishable by a four-year prison sentence. Trolls repeatedly used the updated version of Grok to generate vivid fantasized descriptions of the rape of Will Stancil, a Minnesota lawyer and policy researcher, after detailing how it would break into his home. Stancil stated that he was considering filing a lawsuit against X. Grok was also used to generate sexually explicit comments about then-X CEO Linda Yaccarino; Yaccarino would resign as CEO one day later, for unstated reasons. Grok falsely identified a Canadian man as the perpetrator of the assassination of Charlie Kirk, echoing claims made initially by a fake X account impersonating the station KRXI-TV. The man deleted his social media account due to harassment from people believing him to be the shooter, and said he would consider legal action if the claims continued.

In December 2025, social media users reported that X social network's integrated chatbot, Grok, would allow users to nonconsensually alter photos of individuals, including minors, to show them wearing underwear or bikinis. The majority of these prompts were targeted at women. Images generated by Grok since December 2025 have been disproportionately of people in bikinis, transparent clothes, and the like, with users being able to generate such images through prompts such as "put her in a bikini". This scandal would lead to significant criticism from lawmakers across the world, calls for bans on X, as well as legal crackdowns on X and SpaceXSI for, amongst other reasons, the facilitation of sexual abuse, revenge porn, and child pornography. In August 2026, a plaintiff known as Jane Doe filed a lawsuit alleging that Grok was trained using CSAM of her. When Doe was preschool-aged, she was repeatedly raped by adult men to create CSAM to sell to pedophiles online; the images were hashed by groups including the National Center for Missing and Exploited Children and the Canadian Centre for Child Protection (CCCP). The CCCP notified Doe that it had identified AI CSAM generated by Grok that depicted her, sparking her to file the lawsuit.

In September 2025, Musk announced SpaceXSI was building a new AI-generated online encyclopedia, to be called Grokipedia. The project was suggested and named by David Sacks at the All-In podcast conference earlier that month. Musk described the project as an AI-powered alternative to Wikipedia with fewer biases and inaccuracies. Gizmodo compared the plan to the 2006 Conservapedia project. On October 6, 2025, Musk announced that the early beta version of Grokipedia was scheduled for release later that month. Grokipedia launched on October 27 with over 800,000 articles, compared to Wikipedia's nearly 8 million articles. Some articles are nearly identical to their Wikipedia entries, but the format of Grokipedia citations is different. Wired found Grokipedia's articles to display a right-leaning political bias and some scientific and historical inaccuracies, such as a false claim that pornography contributed to the 1980s AIDS epidemic. In August 2026, Lawfare reported that Grokipedia did not appear to have accepted or rejected suggested edits after April 24, 2026, and that sampled articles had not changed for more than three months.

A chatbot (originally chatterbot) is a software application or web interface designed to converse through text or speech, functioning as a conversation partner. Chatbots have existed for decades, but chatbots based on deep learning have gained popularity during the AI boom of the 2020s, with the releases of generative AI (GenAI) chatbots such as ChatGPT, Gemini, Claude, DeepSeek and Grok. These chatbots typically use fine-tuned large language models (LLMs) to generate text. A major area where chatbots have long been used is customer service and support, with various sorts of virtual assistants.

Modern chatbots like ChatGPT are often based on foundational large language models called generative pre-trained transformers (GPT). They are based on a deep learning architecture called the transformer, which contains artificial neural networks. They generate text after being trained on a large text corpus, exhibiting capabilities beyond their explicit training . Chatbots integrated into apps and websites can call image-generation models or search the web. Some platforms also enable users to interact with conversational interfaces directly through web-based chat environments, allowing real-time assistance, content generation, and task automation without requiring software installation.

3. Dataset Ingestion, Preprocessing & Stochastic Augmentations

In New Zealand, the chatbot SAM – short for Semantic Analysis Machine – has been developed by Nick Gerritsen of Touchtech. It is designed to share its political thoughts, for example on topics such as climate change, healthcare and education, etc. It talks to people through Facebook Messenger. In 2022, the chatbot "Leader Lars" or "Leder Lars" was nominated for The Synthetic Party to run in the Danish parliamentary election, and was built by the artist collective Computer Lars. Leader Lars differed from earlier virtual politicians by leading a political party and by not pretending to be an objective candidate. This chatbot engaged in critical discussions on politics with users from around the world. In India, the state government has launched a chatbot for its Aaple Sarkar platform, which provides conversational access to information regarding public services managed.

Research has identified potential barriers to entry that come with the usage of chatbots for mental health. There are ongoing privacy concerns with sharing user's personal data in chat logs with chatbots. There is a lack of willingness from those in lower socioeconomic statuses to adopt interactions with chatbots as a meaningful way to improve upon mental health. Though chatbots may be capable of detecting simple human emotions in interactions with users, they are incapable of replicating the level of empathy that human therapists do. There is a lack of research about how exactly these interactions help with a user's real life. There are concerns regarding the safety of users when interacting with such chatbots. When improvements and advancements are made to such technologies, how that may affect humans is not a priority. It is possible that this can lead to "unintended negative consequences, such as biases, inadequate and failed responses, and privacy issues". A risk in the usage of chatbots to deal with mental health is increased isolation, as well as a lack of support in times of crisis. A 2025 study by Sentio University evaluated how six major chatbots responded to disclosures of suicide risk and other acute mental health crises, finding that none consistently met clinician determined safety standards. Another notable risk is a general lack of a strong understanding of mental health. Studies have indicated that mental-health-oriented chatbots have been prone to recommending users medical solutions and to rely upon themselves heavily. Obsessive use of chatbots has been linked to chatbot psychosis in people already prone to delusional and conspiratorial thinking. This is caused in part by chatbots "hallucinating" information, as they are designed for engagement, and to keep people talking.

Generative AI uses a high amount of electric power. Due to reliance on fossil fuels in its generation, this increases air pollution, water pollution, and greenhouse gas emissions. In 2023, a question to ChatGPT consumed on average 10 times as much energy as a Google search. Data centres in general, and those used for AI tasks specifically, use significant amounts of water for cooling.

Mathematical Formulation 1.1: Loss Objectives and Optimization Dynamics for Chatbot Projects. The training protocol of chatbot projects minimizes an empirical risk objective defined across high-dimensional parameter space Theta in R^d. Categorical predictions are governed by Cross-Entropy Loss: L_CE = -sum(y_i * log(p_i)). To tackle class imbalance, the network incorporates Focal Loss: L_Focal = -alpha_t * (1 - p_t)^gamma * log(p_t), where gamma modulates the gradient penalty on hard misclassified instances. For spatial tasks, Dice Loss balances continuous spatial overlap: L_Dice = 1 - (2 * sum(y * y_hat) + eps) / (sum(y) + sum(y_hat) + eps). Stochastic optimization utilizes AdamW with decoupled weight decay, maintaining stable moment estimations m_t and v_t across variable learning rate schedules.

Architecture Specification 1.2: Feature Extraction and Layer Formulations for Chatbot Projects. The primary computational backbone for Chatbot Projects incorporates hierarchical deep neural layers designed for multi-scale representation extraction. Input tensors X in R^(B x C x H x W) pass through initial stem convolutions with Batch Normalization and GELU non-linearities. Residual identity skip connections, expressed as F(X) + X, prevent vanishing gradient degradation across deep topologies. Multi-head self-attention mechanisms compute Query, Key, and Value projections: Attention(Q, K, V) = softmax((Q * K^T) / sqrt(d_k)) * V, capturing non-local contextual dependencies across spatial and temporal sequences.

4. Neural Network Backbone, Layer Specifications & Loss Formulations

Data Pipeline 1.3: Stochastic Augmentation and Normalization Protocols for Chatbot Projects. To prevent overfitting and maximize generalization on unseen testing distributions, input datasets undergo rigorous preprocessing. Images and sequences are standardized using z-score normalization: x_norm = (x - mu) / sigma. Stochastic augmentations include random horizontal flips, affine rotations (-15 deg to +15 deg), CutMix, and Mixup: X_tilde = lambda * X_i + (1 - lambda) * X_j, smoothing decision boundaries and reducing model overconfidence.

Deployment Optimization 1.4: TensorRT Acceleration and Edge Quantization for Chatbot Projects. Production deployment of Chatbot Projects focuses on minimizing memory footprints and sub-20ms inference latencies. Trained models are exported to ONNX and compiled via NVIDIA TensorRT. Post-Training Quantization (PTQ) maps FP32 weights into low-bit INT8 representations: q = round(r / S) + Z, achieving high-throughput execution across embedded edge platforms like NVIDIA Jetson without accuracy loss.

Mathematical Formulation 2.1: Loss Objectives and Optimization Dynamics for Chatbot Projects. The training protocol of chatbot projects minimizes an empirical risk objective defined across high-dimensional parameter space Theta in R^d. Categorical predictions are governed by Cross-Entropy Loss: L_CE = -sum(y_i * log(p_i)). To tackle class imbalance, the network incorporates Focal Loss: L_Focal = -alpha_t * (1 - p_t)^gamma * log(p_t), where gamma modulates the gradient penalty on hard misclassified instances. For spatial tasks, Dice Loss balances continuous spatial overlap: L_Dice = 1 - (2 * sum(y * y_hat) + eps) / (sum(y) + sum(y_hat) + eps). Stochastic optimization utilizes AdamW with decoupled weight decay, maintaining stable moment estimations m_t and v_t across variable learning rate schedules.

Architecture Specification 2.2: Feature Extraction and Layer Formulations for Chatbot Projects. The primary computational backbone for Chatbot Projects incorporates hierarchical deep neural layers designed for multi-scale representation extraction. Input tensors X in R^(B x C x H x W) pass through initial stem convolutions with Batch Normalization and GELU non-linearities. Residual identity skip connections, expressed as F(X) + X, prevent vanishing gradient degradation across deep topologies. Multi-head self-attention mechanisms compute Query, Key, and Value projections: Attention(Q, K, V) = softmax((Q * K^T) / sqrt(d_k)) * V, capturing non-local contextual dependencies across spatial and temporal sequences.

Data Pipeline 2.3: Stochastic Augmentation and Normalization Protocols for Chatbot Projects. To prevent overfitting and maximize generalization on unseen testing distributions, input datasets undergo rigorous preprocessing. Images and sequences are standardized using z-score normalization: x_norm = (x - mu) / sigma. Stochastic augmentations include random horizontal flips, affine rotations (-15 deg to +15 deg), CutMix, and Mixup: X_tilde = lambda * X_i + (1 - lambda) * X_j, smoothing decision boundaries and reducing model overconfidence.

5. Training Protocol, Convergence Behavior & Experimental Evaluation

Deployment Optimization 2.4: TensorRT Acceleration and Edge Quantization for Chatbot Projects. Production deployment of Chatbot Projects focuses on minimizing memory footprints and sub-20ms inference latencies. Trained models are exported to ONNX and compiled via NVIDIA TensorRT. Post-Training Quantization (PTQ) maps FP32 weights into low-bit INT8 representations: q = round(r / S) + Z, achieving high-throughput execution across embedded edge platforms like NVIDIA Jetson without accuracy loss.

Mathematical Formulation 3.1: Loss Objectives and Optimization Dynamics for Chatbot Projects. The training protocol of chatbot projects minimizes an empirical risk objective defined across high-dimensional parameter space Theta in R^d. Categorical predictions are governed by Cross-Entropy Loss: L_CE = -sum(y_i * log(p_i)). To tackle class imbalance, the network incorporates Focal Loss: L_Focal = -alpha_t * (1 - p_t)^gamma * log(p_t), where gamma modulates the gradient penalty on hard misclassified instances. For spatial tasks, Dice Loss balances continuous spatial overlap: L_Dice = 1 - (2 * sum(y * y_hat) + eps) / (sum(y) + sum(y_hat) + eps). Stochastic optimization utilizes AdamW with decoupled weight decay, maintaining stable moment estimations m_t and v_t across variable learning rate schedules.

Architecture Specification 3.2: Feature Extraction and Layer Formulations for Chatbot Projects. The primary computational backbone for Chatbot Projects incorporates hierarchical deep neural layers designed for multi-scale representation extraction. Input tensors X in R^(B x C x H x W) pass through initial stem convolutions with Batch Normalization and GELU non-linearities. Residual identity skip connections, expressed as F(X) + X, prevent vanishing gradient degradation across deep topologies. Multi-head self-attention mechanisms compute Query, Key, and Value projections: Attention(Q, K, V) = softmax((Q * K^T) / sqrt(d_k)) * V, capturing non-local contextual dependencies across spatial and temporal sequences.

Data Pipeline 3.3: Stochastic Augmentation and Normalization Protocols for Chatbot Projects. To prevent overfitting and maximize generalization on unseen testing distributions, input datasets undergo rigorous preprocessing. Images and sequences are standardized using z-score normalization: x_norm = (x - mu) / sigma. Stochastic augmentations include random horizontal flips, affine rotations (-15 deg to +15 deg), CutMix, and Mixup: X_tilde = lambda * X_i + (1 - lambda) * X_j, smoothing decision boundaries and reducing model overconfidence.

Deployment Optimization 3.4: TensorRT Acceleration and Edge Quantization for Chatbot Projects. Production deployment of Chatbot Projects focuses on minimizing memory footprints and sub-20ms inference latencies. Trained models are exported to ONNX and compiled via NVIDIA TensorRT. Post-Training Quantization (PTQ) maps FP32 weights into low-bit INT8 representations: q = round(r / S) + Z, achieving high-throughput execution across embedded edge platforms like NVIDIA Jetson without accuracy loss.