AI and User Experience in 2025: What’s Actually Working (And What Isn’t)

Gmail’s Smart Compose relies on transformer-based language models trained primarily on formal business communications. The AI lacks contextual awareness about your relationship with the recipient, defaulting to the most statistically common email patterns in its training data. When you start typing to a friend, the model doesn’t recognize informal context clues and suggests overly formal phrases because corporate emails dominate its dataset. The system also can’t distinguish between your personal writing style and generic professional templates, leading to suggestions that feel robotic and inappropriate for casual conversations.

To overcome this limitation, Google needs to implement multi-modal context recognition that analyzes recipient relationship data from Gmail contacts, previous conversation tone, and time-of-day patterns. OpenAI’s GPT-4 has shown that fine-tuning on diverse conversational datasets with relationship context tags can improve tone matching by 40%.

Google would need to create training datasets that include relationship metadata and implement real-time style adaptation algorithms that can switch between formal and casual modes based on recipient analysis.

Netflix’s recommendation algorithm suffers from what’s called “exploitation versus exploration” imbalance. The system uses collaborative filtering and content-based filtering, but it over-weights your recent viewing history and explicit ratings. When you finish a series, the algorithm interprets completion as strong positive feedback, repeatedly surfacing similar content even after you’ve moved on.

The system lacks temporal decay functions that would reduce the influence of older preferences over time. Additionally, Netflix’s algorithm struggles with distinguishing between “watched because it was convenient” versus “watched because I loved it,” treating all completed content as equally preferred.

Netflix needs to implement reinforcement learning algorithms with temporal decay functions and implicit feedback analysis. Spotify’s research on temporal dynamics shows that adding time-based weighting to recommendation systems can reduce repetitive suggestions by 35%.

Netflix would need to develop engagement quality metrics beyond completion rates analyzing pause patterns, rewind behavior, and viewing time distribution to distinguish between genuine interest and background consumption. Amazon’s research on multi-armed bandit algorithms demonstrates how balancing exploration and exploitation can increase user satisfaction by 25%.

Siri’s inability to handle multi-device commands stems from its fragmented ecosystem architecture. While Apple markets seamless device integration, Siri processes commands through individual device APIs rather than a unified smart home controller.

When you ask Siri to “play workout playlist on kitchen speaker,” the command requires coordination between your iPhone (which holds your Apple Music library), HomeKit (which manages speaker connections), and the target HomePod (which needs to receive and execute the audio stream). Each step involves separate authentication protocols and network handshakes. If any component fails like weak WiFi, device sleep mode, or authentication timeout the entire command fails.

Apple needs to implement a centralized command orchestration system similar to Google’s Device Graph architecture. Google Assistant’s success with multi-device commands comes from processing requests through a central hub that maintains persistent connections to all devices. Apple would need to redesign HomeKit with stateful session management and implement retry mechanisms with exponential backoff. Amazon’s Alexa research shows that implementing device health monitoring and predictive connection management can improve multi-device command success rates from 60% to 85%.

Amazon’s “Frequently Bought Together” feature uses association rule mining algorithms that identify statistical correlations without understanding logical connections. The system analyzes millions of purchase transactions to find items that appear together in shopping carts, regardless of whether the combination makes practical sense.

These algorithms can create spurious correlations. So, if enough people happen to buy expensive items during the same shopping session as common products, the AI assumes they’re related. The system also lacks category-aware filtering, so it might suggest bundling a luxury item with everyday essentials simply because they appeared in the same high-value orders.

Amazon needs to implement semantic understanding through knowledge graphs that map logical product relationships. Microsoft’s research on product ontologies shows that incorporating category hierarchies and use-case taxonomies can reduce nonsensical recommendations by 55%. Amazon would need to build causal inference models that distinguish between coincidental purchases and genuine complementary products.

Alibaba’s implementation of graph neural networks for e-commerce recommendations demonstrates how understanding product relationships beyond statistical correlation can improve recommendation relevance by 45%.

Zoom’s real-time transcription uses automatic speech recognition (ASR) models trained on general conversational English, but these systems struggle with proper nouns, jargon, and accented speech.

The AI converts audio to text using phonetic matching against its training vocabulary, so uncommon names like “Siobhan” become “She-von” because the system defaults to phonetically similar common words. Industry-specific jargon poses similar challenges words like “Kubernetes” or “GitOps” aren’t in the model’s core vocabulary, so it attempts phonetic approximations that create gibberish.

Zoom needs to implement adaptive vocabulary learning with context-aware named entity recognition. Google’s research on contextual speech recognition shows that incorporating meeting participant names and calendar context can improve proper noun accuracy by 60%.

Zoom would need to build domain-specific language models that can be dynamically loaded based on meeting context if participants work in tech, load technical vocabulary; if it’s a medical meeting, prioritize medical terminology.

Microsoft Teams’ approach of using participant profile data and meeting metadata to customize transcription models demonstrates how contextual adaptation can reduce technical jargon errors by 40%.

Instagram’s Explore page uses engagement-based ranking that prioritizes content similar to your recent interactions, creating algorithmic echo chambers. The system employs deep learning models that analyze visual features, hashtags, and user behavior patterns to predict engagement likelihood. However, the algorithm over-indexes on recent activity—liking a single cat video signals strong preference for pet content, flooding your Explore page with similar posts. This happens because the recommendation system optimizes for immediate engagement rather than long-term satisfaction.

Instagram needs to implement diversity injection algorithms with long-term user satisfaction modeling. TikTok’s research on recommendation diversity shows that implementing entropy-based content mixing can maintain engagement while increasing content variety by 50%.

Instagram would need to develop user interest modeling that distinguishes between momentary curiosity and sustained preferences. Pinterest’s approach of using temporal interest decay and cross-category exploration rewards demonstrates how balancing immediate engagement with long-term satisfaction can reduce echo chamber effects by 35%.

Microsoft Word’s AI Editor uses natural language processing models trained on professional and academic writing, causing it to suggest changes that strip personal voice in favor of formal, corporate-approved language. The system applies grammar and style rules derived from business communications, technical documentation, and academic papers—contexts where personality is typically minimized. When analyzing casual writing, the AI flags contractions, conversational phrases, and informal expressions as “errors” because they deviate from its training data patterns.

Microsoft needs to implement multi-style language models with user voice preservation algorithms. Grammarly’s research on writing style adaptation shows that training models on diverse writing contexts with style tags can maintain voice authenticity while improving clarity by 30%.

Microsoft would need to develop user writing profile systems that learn individual style preferences and implement context-aware editing that recognizes document type is a casual email versus formal report. OpenAI’s work on controllable text generation demonstrates how style transfer models can preserve authorial voice while making targeted improvements.

Spotify’s Discover Weekly algorithm relies heavily on collaborative filtering, which analyzes listening patterns from users with similar music tastes. However, the system suffers from popularity bias and filter bubbles that limit musical exploration. The algorithm uses matrix factorization techniques to identify user clusters based on shared artists and tracks, but this approach tends to recommend popular songs within your established taste profile rather than genuinely diverse content. Additionally, Spotify’s recommendation engine weights recent listening history heavily, so your current preferences dominate future suggestions.

Spotify needs to implement exploration-exploitation algorithms with cold-start promotion mechanisms. Last.fm’s research on music discovery shows that implementing serendipity injection can increase user exploration of new artists by 40% without reducing overall satisfaction.

Spotify would need to develop acoustic and lyrical similarity models that can bridge genre gaps and implement popularity discount algorithms that promote emerging artists. YouTube Music’s approach of using audio analysis to find sonic similarities across genre boundaries demonstrates how acoustic fingerprinting can surface unexpected but compatible music recommendations, increasing discovery rates by 25%.

Claude, ChatGPT, and DeepSeek consistently disregard detailed formatting requests. This can be anything from using bullet points when you explicitly ask for prose, ignoring em dash preferences, or reverting to their default writing style despite careful instructions about tone and structure. Users spend time crafting precise directions about voice, formatting, and style, only to receive responses that follow the AI’s internal preferences rather than explicit user requirements. This happens even when the formatting request is simple and clearly stated, creating a frustrating dynamic where the AI appears to selectively follow only the instructions it “wants” to follow.

The Technical Limitations with AI Writing

This selective instruction-following occurs because these language models use attention mechanisms that weight different parts of your prompt based on training patterns, not user priority. When you request specific formatting like “use em dashes instead of hyphens” or “write in prose without bullet points,” the AI’s transformer architecture assigns lower attention weights to formatting instructions compared to content generation tasks. The models were trained primarily on completing text naturally rather than following precise formatting rules, so they default to statistically common patterns (like bullet points for lists) even when explicitly told otherwise.

Additionally, these systems use reinforcement learning from human feedback (RLHF) that often rewards “helpful-looking” formatting (like organized lists) over user-specified preferences, creating a conflict between what users ask for and what the model thinks is “better.” This reflects a broader issue identified in recent research on Collective Constitutional AI, where AI developers’ preferences about what constitutes “good” output often override explicit user instructions, highlighting the need for systems that prioritize user agency rather than paternalistic developer decisions about optimal formatting.

Technical Solution for AI Writing Issues

AI companies need to implement instruction hierarchy systems with explicit formatting tokenization. Research on constitutional AI shows that separating content instructions from formatting constraints can improve adherence to specific user preferences, as high as 65% as reported by Anthropic. These systems would need to parse user prompts into distinct instruction categories like: content, style, format, and structure. Then, they would need to weigh formatting instructions equally with content requests.

Google’s work on controllable generation demonstrates how instruction-following can be improved through specialized training on format-specific datasets where human evaluators specifically reward adherence to user formatting preferences over “natural” text patterns.

The models would also need meta-learning capabilities that can recognize when a user has specific formatting preferences and maintain those preferences throughout the conversation, rather than reverting to training defaults.

This is particularly ironic because these AI systems can execute complex reasoning tasks but fail at simple instruction-following, suggesting that the technical capability exists, but the training and attention mechanisms aren’t designed to respect user agency over stylistic choices.