Midjourney acquired the astrology app Co-Star

Midjourney's acquisition of the astrology app Co-Star seems, at first glance, like a bizarre marriage of generative AI hype and celestial pseudoscience. But for engineers and platform strategists, this move reveals a deliberate bet on something far more valuable than horoscope text: a deeply structured, user-validated dataset of human personality and behavioral cues. Midjourney's purchase of Co-Star signals a shift from generative media to personalized AI experiences that mine intimate user data. This isn't about reading star charts-it's about training models that know you better than you know yourself.

Co-Star, launched in 2017, built a loyal Gen Z and millennial following by generating algorithmically crafted horoscopes based on precise birth date, time. And location. The app's appeal lies in its hyper‑personalized, often brutally honest predictions-delivered in a distinct, literary tone. Behind the scenes, Co-Star runs a sophisticated rule engine that blends astronomical calculations (ephemeris data) with natural language generation (NLG) templates. For Midjourney, this acquisition isn't about astrology; it's about acquiring a behavioral identity layer that can be fed into diffusion models - recommendation engines. And even future conversational agents.

Let's examine what this means from a technical, strategic. And ethical standpoint-and why every engineer building consumer-facing AI should pay close attention.

The Strategic Logic: Why an AI Lab Sees Gold in Astrology Data

Midjourney's core product is text‑to‑image generation. The company's competitive advantage has been its community‑driven model fine‑tuning and aesthetic quality. But generation alone is a commoditizing market. To differentiate, Midjourney needs personalized generation-outputs that adapt to individual users' tastes, moods,, and and identityCo-Star provides a ready‑made taxonomy: the 12 zodiac signs - 10 planets, houses, aspects-each mapping to a rich set of personality descriptors and archetypes. This is far more structured than freeform user profiles.

Consider the engineering challenge: generating an image that resonates with a "Leo with a Virgo rising" requires mapping astrological attributes to latent space vectors. Midjourney could fine‑tune a diffusion model using Co-Star's existing data (e g., user demographics, birth‑chart labels, and engagement patterns). The result would be a model that not only understands style but also personality‑driven aesthetics. For example, a "fire sign" prompt might yield warmer colors and dynamic compositions, while an "earth sign" prompt leans toward grounded, natural scenes.

In production environments, we have seen similar personalization attempts using generic survey data or cookie tracking. But those lack the symbolic richness and self‑identification that astrology provides. Co-Star's users voluntarily share birth data because they believe it explains their identity-a level of engagement that synthetic or scraped data rarely matches.

What the Co-Star Dataset Offers That Synthetic Data Cannot

Co-Star's backend collects precise birth coordinates (date, time, location), which are then used to calculate a full natal chart via libraries like Swiss Ephemeris. Beyond that, the app logs user reactions: swipes, saves, shares, and direct feedback ("this is so true"). This creates a longitudinal dataset of self‑reported personality validation-users explicitly confirm when a generated text matches their self‑image. That ground truth feedback is incredibly rare and valuable.

Synthetic data, no matter how well generated, lacks this trusted signal. You can't synthetically produce a user's affirmation that a model's output aligns with their lived identity. For Midjourney, this dataset enables a new class of training: reinforcement learning from human feedback (RLHF) on personality‑matched outputs. Instead of just aesthetic quality scores, the reward function can now incorporate personal relevance,

This raises immediate data privacy concernsBirth data is classified as personally identifiable information (PII) under GDPR and CCPA-especially when combined with precise geolocation. Midjourney must add robust anonymization and consent management. Co-Star has a history of API integrations (e g., syncing with Facebook), which have drawn scrutiny, but midjourney's next steps should include a transparent data usage policy and opt‑in model for any personalization features.

From Image Generation to Personalized Content Pipelines

Midjourney currently operates as a standalone image generation service. But the Co-Star acquisition hints at a broader platform vision. Imagine a future where a user authenticates with their Co-Star profile and receives not only images but also personalized narratives, curated music playlists. Or even journaling prompts-all generated by Midjourney's models and tailored to their astrological profile.

Technically, this requires building a personalization layer between the user and the generative model. This layer would convert astrological attributes into prompt modifiers: incorporating sign‑specific keywords, color palettes. And composition rules. For example, a "Pisces" prompt could trigger dreamlike, water‑based imagery. While "Aries" calls for bold, fiery elements. The model itself doesn't need to be retrained from scratch; instead, lightweight embeddings (like LoRA adapters) can be applied per user or per sign group.

This approach mirrors how large language models (LLMs) are being fine‑tuned for specific personas (e g., a "friendly assistant" vs. And a "sarcastic chatbot")Midjourney could release an API that accepts a birth chart as input and returns a "persona embedding" that can be appended to any generation request. Developers integrating Midjourney into their apps would suddenly have access to hyper‑personalized generation without managing the astronomical calculations themselves.

The Broader Trend: AI Labs Acquiring Data-Rich Consumer Apps

Midjourney isn't alone in this strategy. We have seen AI companies acquire apps with unique, engaged user bases to fuel data‑hungry models. For example, Character. AI's acquisition by Google (rumored) gave access to massive conversational data. Microsoft's LinkedIn acquisition was about professional identity data. Midjourney's move focuses on personal identity-a different, more emotional dataset.

This trend will accelerate as generative models move beyond "one‑size‑fits‑all" outputs, and personalized generation requires personalized data,And consumer apps that solicit intimate user information (horoscopes, personality tests, mood trackers) become prime acquisition targets. The engineering challenge is integrating disparate data formats and ensuring low‑latency personalization at scale. Midjourney's infrastructure (likely custom GPU clusters with fast inference) must now also handle real‑time birth chart calculations-a well‑understood but compute‑cheap task.

We should also expect competition from other labs: Stability AI might acquire a journaling app; OpenAI might buy a Myers‑Briggs service. The data land grab is shifting from web scrapes to self‑reported human traits.

Technical Architecture Considerations for Integration

Co-Star's backend is built on Node js and possibly Python for ephemeris calculations. Midjourney's infrastructure is proprietary but known to use large‑scale inference queues and custom diffusion models. Merging these two stacks presents several engineering hurdles:

• Data pipeline latency: Co-Star generates horoscopes on a schedule (daily) with burst traffic. Midjourney's inference is near‑real‑time. The combined system must support both batch and on‑demand personalization.
• Prompt engineering at scale: Converting astrological scores into effective image prompts requires experimentation. A/B testing infrastructure is needed to measure user satisfaction (e g, and, "did you like this image").
• Model fine‑tuning cadence: Co-Star's data grows continuously. Midjourney will need to add automated retraining pipelines (e g., using MLOps tools like MLflow) to incorporate new user feedback without compromising base model quality.
• Privacy‑preserving computation: Birth data shouldn't be sent to the model servers in plain text. Techniques like homomorphic encryption or differential privacy might be considered,, and though they add latencyA pragmatic approach is on‑device conversion of birth data to a sign‑based token (e. And g, "aries‑gemini‑ascendant") that retains utility without exposing raw PII.

These are non‑trivial but solvable engineering problems. Midjourney's team, known for its lean and agile development culture (the company has fewer than 50 employees as of early 2025), will need to grow or partner with infrastructure specialists to execute the integration.

User Privacy and Ethical Boundaries in Astrological AI

Astrology is pseudoscience. But users' belief in it's real-and so are the privacy risks. Co-Star previously faced criticism for sharing user data with Facebook (via the now‑restricted Graph API). Midjourney must tread carefully to avoid backlash. Key considerations:

• Explicit consent for model training: Users should be asked whether their birth data can be used to improve Midjourney's models, with the ability to opt out without losing access to the app.
• Data minimization: Only relevant attributes (e, and g, zodiac sign, not precise birth coordinates) should be transmitted to inference servers.
• Transparency about personalization: Users should be told when a generated image is based on their astrological profile. And given the option to disable it.
• Bias and fairness: Astrology is culturally specific. Midjourney's models might inadvertently reinforce stereotypes associated with certain signs. Which could offend users outside the target demographic.

From a compliance standpoint, the acquisition will likely trigger privacy audits under GDPR and CCPA. Midjourney's prior data practices (they collect user prompts and generated images) are relatively straightforward. Adding birth data introduces a new risk category that their legal and engineering teams must address proactively.

Impact on the Developer Ecosystem and Open Source Alternatives

Midjourney's closed‑source model means developers currently interact only through the Discord bot or API. The Co-Star integration could be exposed as an optional personalization service within the API, allowing third‑party apps to pass astrological parameters for tailored generations. This would be a premium feature, likely increasing revenue per user.

For open‑source enthusiasts, the Ephemeris calculation itself is already open (e, and g, Python's pyephem library). However, Co-Star's proprietary mapping of planetary aspects to personality traits-and its tuned NLG templates-will remain behind closed doors. Developers can add similar personalization by building their own trait vectors from user surveys, but they will lack the scale of Co-Star's validated dataset.

This acquisition may also spur the creation of open‑source "personality‑to‑prompt" libraries, much like the Stable Diffusion